SaaS Glossary — Plain-English Definitions
158 plain-English definitions covering SaaS metrics, AI concepts, proxy types, web scraping, and affiliate marketing — written for practitioners, not textbooks.
158 of 158 terms
3
3PL (Third-Party Logistics)
A third-party logistics provider stores your inventory, picks and packs orders, and ships them to customers on your behalf. You send stock to their warehouse; they handle everything from the shelf to the doorstep, including returns. Ecommerce brands use a 3PL to avoid signing warehouse leases and hiring packing staff, and to place inventory near customers so delivery is faster and cheaper. The trade-off is losing direct control over packaging quality and fulfilment speed, and paying per-order fees plus storage that can erode margin at low volume.
A
AI Agent
An AI agent is an LLM-based system that can take autonomous actions — calling tools, browsing the web, writing code, querying databases — to complete a multi-step task without manual intervention at each step. Unlike a chatbot that generates text, an agent executes work in the world. Agents use a perceive-reason-act loop: read the environment, decide what to do, call a tool or write output, observe the result, and repeat. The risk is that errors compound across steps, making robust error handling, sandboxing, and human oversight critical for production deployments. Claude, GPT-4, and Gemini all support function calling as the foundation for agent behavior.
API Key
A unique string that identifies and authenticates your application when it calls an API. The server checks the key on every request to confirm who you are, apply your rate limits, and bill your usage. Treat an API key like a password: never commit it to a public repository or embed it in front-end JavaScript, where anyone can read it and run up your bill. Best practice is to store keys in environment variables or a secrets manager, scope them to the minimum permissions needed, and rotate them periodically or immediately after any suspected leak.
API Rate Limit
A cap on how many API requests you may make in a given window — for example 100 requests per minute or 10,000 per day. Providers enforce limits to protect infrastructure and to separate pricing tiers. Exceeding one usually returns HTTP 429 (Too Many Requests), often with a Retry-After header telling you how long to wait. Robust integrations handle this with exponential backoff (waiting progressively longer between retries) rather than hammering the endpoint, and by batching or caching requests so the same data is not fetched repeatedly.
API-First
API-first means the product is designed so every feature is accessible via API before — or simultaneously with — any UI. Developers can build on top of it programmatically, making it composable within other stacks and workflows. The API is the product, not an afterthought. API-first products dominate developer infrastructure: Stripe for payments, Twilio for communications, Cloudflare for networking. The approach unlocks a second distribution channel — developer adoption — and a powerful moat, because switching requires unwinding embedded API calls across an entire codebase, not just a dashboard login.
ARPU (Average Revenue Per User)
ARPU is total recurring revenue divided by the total number of paying accounts for a given period. It tells you the average deal size across your customer base and is the baseline for LTV calculations. Rising ARPU usually signals successful upsell motions or healthy expansion — but can also mask churn of small accounts. Falling ARPU is often the first signal that a product is attracting a different, lower-value customer segment than intended. ARPU segmented by cohort or plan tier is far more useful than blended ARPU for diagnosing pricing strategy.
ARR (Annual Recurring Revenue)
Annual Recurring Revenue is MRR × 12 — the yearly run-rate of all active subscription revenue at a point in time. It excludes one-time fees, services, and overages. ARR is the headline metric investors and boards track because it smooths monthly noise and maps cleanly to the annual contract cycle most SaaS companies sell.
Activation Rate
The share of new signups who reach a defined 'aha' moment — the first action that reliably predicts they will stick around. That milestone is product-specific: sending a first campaign, inviting a teammate, or connecting a data source. Activation matters more than raw signup volume because users who never activate almost never convert to paid and rarely return. Improving it usually means shortening the path to first value: better onboarding, sensible defaults, templates, or removing setup steps that can wait until later.
Affiliate Network
An affiliate network is a marketplace that connects advertisers (merchants) with publishers (affiliates). The network handles tracking, attribution, fraud detection, and payments — taking a 20-30% cut of commission volume. Well-known B2B SaaS networks include PartnerStack, Impact, and Tune. Consumer networks include CJ, Awin, ShareASale, and Rakuten. Choosing a network is mostly about which merchants they already represent in your vertical.
Agentic AI
AI systems that pursue a goal over multiple steps rather than answering a single prompt. An agent plans, calls tools or APIs, observes the result, and adjusts — looping until the task is done or it hits a limit. Where a chatbot answers 'what should I do?', an agent does it: researching, updating a CRM record, or drafting and filing a report. The practical trade-offs are reliability and control, which is why production deployments add guardrails, permission scopes, spend caps and human approval steps before consequential actions.
Annual Contract Value (ACV)
The average annualised revenue from a single customer contract, excluding one-off fees. A three-year deal worth $150,000 in total has an ACV of $50,000. ACV lets you compare contracts of different lengths on equal terms and is a core input to sales planning: it shapes how much you can afford to spend acquiring a customer and which sales motion makes sense. Low ACV demands self-serve or product-led growth, while high ACV can justify field sales, onboarding services and dedicated account management.
Annual Run Rate
Current revenue extrapolated over a full year — typically the latest month's recurring revenue multiplied by twelve. ARR gives a fast, forward-looking read on business size without waiting for audited annual figures, which is why investors and boards lean on it. The caveat is that it assumes the latest period is representative: a seasonal spike, a large one-off contract, or a month of unusual churn will distort it. For subscription businesses it should be built from recurring revenue only, never from one-time services or setup fees.
Anti-Bot Detection
Anti-bot detection is a suite of techniques websites use to identify and block automated requests. It combines IP reputation scoring, behavioral analysis (mouse movement, click patterns, timing), browser fingerprinting, JavaScript challenges, and ML-based anomaly detection into a risk score per request. Enterprise anti-bot systems like Cloudflare Bot Management, Akamai Bot Manager, and DataDome combine dozens of signals to score requests in real time. Bypassing them requires matching the behavioral and technical fingerprint of a real browser — not just spoofing a user agent. The sophistication arms race between scraper tooling and anti-bot vendors drives most of the complexity in modern professional scraping.
Anti-Detect Browser
A browser that masks or randomises the fingerprint your device broadcasts — user agent, screen resolution, installed fonts, canvas and WebGL rendering, timezone and language — so multiple sessions look like different people on different machines. Combined with per-profile proxies, it lets operators run many accounts without them being linked by a shared fingerprint. Common in affiliate marketing, ad verification, ecommerce and web scraping. Note that most platforms' terms of service restrict multi-accounting, so the compliance risk sits with the operator.
Attribution Window
The attribution window is the length of time between a user click and a conversion during which an affiliate or ad source can claim credit. Typical values: 7 days (DTC), 30 days (SaaS), 90 days (high-consideration B2B). Shortening the window improves last-click advertiser economics; lengthening it rewards top-of-funnel partners. Every advertiser fight with their affiliate program eventually becomes a fight about the attribution window.
Average Order Value (AOV)
Total revenue divided by number of orders over the same period. If 200 orders generate $9,000, AOV is $45. It is one of only three ways to grow ecommerce revenue — more visitors, better conversion, or higher AOV — and often the cheapest to move because it requires no extra traffic spend. Typical levers include bundles, volume discounts, free-shipping thresholds set just above current AOV, and post-purchase upsells. Track it alongside margin, since discount-driven AOV gains can quietly reduce profit per order.
B
Bandwidth (Proxy)
In proxy pricing, the volume of data passed through the proxy, billed per gigabyte. Every page you request consumes bandwidth — HTML, images, scripts, fonts and video all count — so a media-heavy site burns far more GB than a text page. Residential and mobile proxies are typically sold this way because their IPs are scarce, while datacenter proxies are more often sold per IP with unlimited traffic. You can cut consumption substantially by blocking images and media in your scraper, requesting only the resources you actually parse.
Bookings
The total contracted value a customer commits to, recorded when the contract is signed rather than when cash arrives or revenue is recognised. A customer signing a $60,000 annual contract creates $60,000 in bookings on day one, even though revenue is recognised monthly and payment may be quarterly. Bookings measure sales momentum, revenue measures delivery, and cash measures collection — three different numbers that routinely diverge. Confusing them is one of the most common causes of SaaS reporting disputes.
Bounce Rate
The percentage of visitors who land on a page and leave without any further interaction. High bounce is not automatically bad: a visitor who reads a glossary definition, gets their answer and leaves has been served well. It matters most on pages designed to drive a next step, such as landing pages and product pages, where bounce signals a mismatch between the promise that brought someone there and what they found. Diagnose it alongside time on page and scroll depth rather than in isolation.
Browser Fingerprinting
Browser fingerprinting is the collection and analysis of browser attributes — installed fonts, screen resolution, WebGL renderer, audio context, timezone, plugin list, canvas rendering output — to create a unique identifier for a client without using cookies. Anti-bot systems use fingerprinting to track clients across IP changes and detect bots that rotate proxies but maintain a consistent technical fingerprint. Defeating fingerprinting requires a full browser environment with realistic, randomized attribute values. Playwright's stealth plugins and commercial tools like Bright Data's Scraping Browser provide hardened browser instances specifically designed to pass fingerprint-based challenges.
Burn Rate
How much cash a company consumes each month. Gross burn is total monthly spend; net burn is spend minus revenue, and is the number that actually matters because it determines how long the bank balance lasts. A startup with $600,000 in the bank and $50,000 net monthly burn has twelve months of runway. Founders track it weekly because it governs every other decision — hiring pace, marketing spend and fundraising timing — and because burn tends to creep upward far more easily than it comes down.
C
CAC (Customer Acquisition Cost)
CAC is the total cost of acquiring one new paying customer — marketing spend, sales salaries, tools, and overhead — divided by the number of customers landed in the same period. It is the single most important cost metric in SaaS because it determines whether growth is profitable. Keeping CAC low relative to LTV is the core unit-economics test. A CAC:LTV ratio of 1:3 or better is considered healthy; below 1:2 usually signals growth is too expensive to sustain. Tracking CAC separately by channel (paid search, outbound, content) reveals which acquisition paths are worth scaling.
CAC Payback Period
Payback period is the number of months required to recover the cost of acquiring a customer through that customer's gross profit. Formula: CAC ÷ (ARPU × gross margin). It answers the question: how long before this customer is profitable? Under 12 months is healthy for SMB SaaS. Enterprise SaaS can tolerate 18–24 months because of higher LTV and lower churn. Payback period directly determines how much capital a company needs to scale — a 6-month payback business can fund growth from revenue far sooner than a 24-month one.
CAPTCHA
CAPTCHA (Completely Automated Public Turing test to tell Computers and Humans Apart) is a challenge mechanism designed to block automated requests. Common forms: image recognition puzzles, audio challenges, invisible behavioral analysis (reCAPTCHA v3), and proof-of-work challenges like Cloudflare's Turnstile. CAPTCHAs are the primary defense layer for high-value scraping targets. Solving them programmatically requires either human-labor CAPTCHA farms, AI-based solver APIs (2captcha, Anti-Captcha, CapSolver), or browser fingerprint quality good enough to pass behavior-based challenges without triggering them. CAPTCHA solving adds latency and cost to any scraping pipeline that encounters it.
CDN (Content Delivery Network)
A globally distributed network of servers that caches your static files — images, CSS, JavaScript, video — and serves each visitor from the location nearest to them. Instead of every request travelling to a single origin server, a user in Sydney is served from an Australian edge node, cutting latency dramatically. CDNs also absorb traffic spikes and DDoS attacks that would overwhelm a single origin, and reduce origin bandwidth costs. They are effectively mandatory for any site with an international audience or heavy media.
CPA (Cost Per Action)
CPA is an affiliate commission model where the advertiser pays only when a user completes a specific action — typically a purchase, subscription signup, lead form submission, or app install. The affiliate bears all traffic-acquisition risk; the advertiser pays only for verified results. CPA is the dominant model in performance affiliate marketing because it aligns incentives cleanly: publishers optimize for quality conversions, not just clicks. Common CPA rates range from $2–$5 for lead generation to 20–40% of the first payment for SaaS subscriptions. Networks like Commission Junction, Impact, and PartnerStack facilitate CPA tracking between advertisers and publishers.
CPC (Cost Per Click)
CPC is a pricing model where the advertiser pays for each click on their ad or affiliate link, regardless of whether the click converts into a purchase or signup. Google Ads and most paid search platforms default to CPC bidding because it ties cost directly to traffic intent signals. In affiliate marketing, CPC deals are less common than CPA because the advertiser bears conversion risk. They appear in content networks, sponsored content placements, and comparison sites where traffic volume matters more than direct attribution. CPC rates vary widely by niche — competitive SaaS categories can run $10–$50+ per click on branded terms.
CPM (Cost Per Mille)
CPM is an advertising pricing model where the advertiser pays per 1,000 impressions (views), regardless of clicks or conversions. "Mille" is Latin for thousand. CPM is the standard model for brand awareness campaigns, display advertising, and newsletter sponsorships. For affiliate marketers running display placements, CPM is a useful metric to benchmark media efficiency — but conversion tracking downstream is essential because a high CPM can be justified by strong conversion rates. Newsletter CPMs in B2B SaaS niches typically run $40–$150 per thousand, significantly above broad consumer audiences.
CSS Selector
A CSS selector is a pattern used to identify and extract specific elements from an HTML document, based on element type, class, ID, attribute, or hierarchy. Scrapers use them — via libraries like BeautifulSoup, Cheerio, or Playwright — to pull specific fields from page markup. CSS selectors are often more stable than XPath for scraping because they are tied to visual presentation rather than exact document hierarchy. When a site redesigns its HTML structure, selectors break — building robust selectors that target semantic attributes (data-* attributes, aria roles) rather than class names reduces maintenance burden because semantic attributes tend to be more stable than presentational class names.
Cart Abandonment
When a shopper adds items to their cart but leaves without completing checkout. Industry-wide it affects roughly two-thirds to three-quarters of carts, making it one of the largest recoverable revenue leaks in ecommerce. The most common causes are unexpected shipping costs revealed late, forced account creation, a long or confusing checkout, and limited payment options. Fixes fall into two buckets: removing friction in checkout itself, and recovering intent afterwards through abandoned-cart emails, SMS or retargeting ads.
Cashback
A rebate that returns part of a purchase price to the buyer after the transaction completes. In affiliate marketing, a cashback site shares its commission with the shopper, which is why cashback offers convert well even without a merchant discount. For merchants, cashback partners drive volume but attract price-sensitive buyers and can capture credit for sales that would have happened anyway — a phenomenon known as commission leakage, usually managed through attribution rules and tighter partner terms.
Chain of Thought
A prompting technique in which a model is asked to work through its reasoning step by step before giving a final answer. Making the intermediate steps explicit improves accuracy on multi-step problems — arithmetic, logic, planning — because the model conditions each step on the previous one rather than jumping straight to a conclusion. Modern reasoning models do this internally by default. The costs are more output tokens, higher latency and greater expense, so it is best reserved for genuinely hard tasks.
Chargeback
A forced payment reversal initiated by the cardholder through their bank rather than by asking the merchant for a refund. The merchant loses the revenue, usually pays a fee on top, and must submit evidence to contest it. Card networks monitor chargeback ratios and can impose penalties or terminate processing for merchants who exceed thresholds, which makes prevention critical: clear billing descriptors, obvious cancellation paths, responsive support and delivery confirmation all reduce disputes before they escalate.
Churn
Churn is the rate at which subscription customers cancel or downgrade in a given period. Gross churn counts only departures; net churn subtracts expansion from the same cohort. Healthy B2B SaaS typically sees 0.5-1.5% monthly gross churn. SMB SaaS runs 2-5%. Anything north of 5% signals product-market-fit or pricing issues that compound into revenue death spirals if not addressed.
Click-Through Rate (CTR)
The percentage of people who click something after seeing it — clicks divided by impressions. It is the standard measure of whether a piece of creative earns attention, used across ads, email subject lines and search results. What counts as good varies enormously by context, so CTR is only meaningful compared against your own baseline for the same placement. Beware optimising it alone: a misleading headline can lift CTR while lowering conversion and increasing refunds, making the campaign worse overall.
Cohort Analysis
Grouping users by a shared starting characteristic — usually their signup month — and tracking each group's behaviour over time instead of blending everyone into one average. This separates the health of the product from the effect of growth. Aggregate retention can look stable purely because new signups mask older users leaving; cohort curves expose that immediately. It is also the only reliable way to tell whether product changes actually improved retention, because you can compare later cohorts against earlier ones.
Cold Outreach
Contacting prospects who have had no prior relationship with you, typically by email, LinkedIn or phone. Effectiveness depends far more on targeting and relevance than on volume: a narrowly defined ideal customer profile with a message referencing something specific about the recipient outperforms mass sending by a wide margin. It is also legally regulated — GDPR in Europe, CAN-SPAM in the US and similar laws elsewhere impose requirements on consent, sender identification and opt-out, with real penalties for ignoring them.
Concurrency (Scraping)
The number of requests a scraper runs simultaneously. Higher concurrency finishes large jobs faster but increases the chance of rate limiting, IP bans and CAPTCHAs, and can degrade the target site for real users. Sustainable scraping tunes concurrency against the target's tolerance, spreads requests across a proxy pool so no single IP is conspicuous, and adds randomised delays between requests. Many providers also cap concurrent connections per plan, making it a pricing dimension as well as a technical one.
Concurrent Connections
Concurrent connections (or concurrency) is the number of simultaneous requests a proxy plan supports at once. Higher concurrency means faster scraping — you can parallelize across many IPs simultaneously rather than waiting for each request to complete before firing the next. Concurrency limits are a key pricing differentiator among proxy providers. Entry plans might allow 10–50 concurrent connections; enterprise plans can run thousands. For large-scale scraping pipelines, practical throughput is usually concurrency × request speed — raising concurrency is often the fastest way to scale a working scraper.
Context Window
The context window is the maximum amount of text — measured in tokens — that an LLM can process in a single request, including both the prompt you send and the response it generates. Claude 3's context window is up to 200K tokens; GPT-4 Turbo supports 128K. The context window determines how much history, documentation, or data a model can "see" at once. Longer contexts enable richer RAG pipelines and multi-document synthesis but also increase latency and cost. Context window limitations drove the development of chunking strategies and vector-search retrieval systems as ways to work around early limits.
Contraction MRR
Contraction MRR is the monthly recurring revenue lost when existing customers downgrade their plan, reduce seats, or move to a lower tier — without cancelling entirely. It is a leading indicator of churn: accounts that downgrade often cancel in the following quarter. High contraction is an early warning sign that customers are trimming budgets, hitting usage limits they do not need, or finding value in only part of the product. Customer success teams use contraction signals to trigger proactive outreach before an account goes dark entirely.
Conversion Funnel
The sequence of steps a visitor passes through on the way to a desired action — for example visit, signup, activation, purchase. Each step loses some proportion of people, so mapping the funnel shows exactly where users drop out and how much revenue each leak costs. The practical value is prioritisation: fixing the stage with the steepest drop-off and the highest traffic almost always beats optimising a later stage that only a fraction of users reach.
Conversion Pixel
A small snippet of tracking code placed on a confirmation or thank-you page that fires when a visitor completes a valuable action, reporting it back to the ad platform. This closes the loop between ad spend and outcome, letting the platform attribute conversions and — more importantly — train its bidding algorithms on who actually converts. Browser privacy changes and ad blockers have made pixel-only tracking increasingly unreliable, which is why many advertisers now pair pixels with server-to-server tracking.
Conversion Rate
The percentage of visitors who complete a target action — buying, signing up, booking a demo — out of everyone who had the chance. It is the efficiency multiplier on all your traffic: doubling conversion has the same revenue effect as doubling visitors, usually at far lower cost. Always define the denominator explicitly, because 'conversion rate' means different things measured per session, per unique visitor or per qualified lead, and comparing figures across those definitions produces nonsense.
Coupon Stacking
Applying more than one discount to a single order — for example a percentage-off code plus free shipping plus store credit. Most ecommerce platforms disallow it by default because stacked discounts can wipe out margin or even produce negative-value orders. Merchants who allow it deliberately usually cap the combined discount, restrict which codes combine, or exclude sale items. Shoppers should assume codes will not stack unless a merchant explicitly says otherwise.
Customer Health Score
A composite metric that predicts whether an account is likely to renew, churn or expand, built from signals such as product usage depth, feature adoption, support ticket volume and sentiment, and stakeholder engagement. The point is to make churn visible early enough to act, rather than discovering it at renewal. Scores are only as good as their inputs, so most teams calibrate them against actual churn outcomes and revise the weightings once real patterns emerge.
D
DOM (Document Object Model)
The DOM is the in-memory tree representation of an HTML document that the browser builds after parsing HTML and executing JavaScript. It is the live structure that JavaScript manipulates — and the structure that browser-based scrapers like Playwright and Puppeteer interact with. Static scrapers read raw HTML without executing JavaScript. Browser-based scrapers interact with the fully rendered DOM — necessary for JavaScript-heavy single-page apps where content only appears after client-side rendering. DOM-based scraping is slower and more resource-intensive but handles modern web apps that a raw HTTP scraper would see as nearly empty.
Data Parsing
Extracting structured fields from unstructured or semi-structured source data — pulling a price, title and rating out of a scraped HTML page and turning them into database columns. Parsing is usually the most brittle part of any scraping pipeline, because a small change to the source site's markup can silently break your selectors. Resilient parsers prefer stable attributes over fragile CSS paths, validate the shape of what they extract, and alert on sudden drops in extraction rate rather than failing quietly.
Datacenter Proxy
A datacenter proxy routes traffic through an IP address hosted in a commercial hosting provider — AWS, Hetzner, OVH, DigitalOcean — rather than a residential ISP. Datacenter IPs are fast, cheap, and highly detectable. Use them for public-data scraping, CI/CD egress, geo-testing, and anything where the target site does not actively fingerprint IP reputation. Expect per-GB pricing in the $0.10-$0.50 range versus $2-$8 for residential.
Deep Linking (Affiliate)
Deep linking in affiliate marketing means generating a tracked link that sends users directly to a specific product page, landing page, or pricing page — rather than the advertiser's homepage. Deep links improve conversion by reducing friction between ad click and purchase intent. Most affiliate networks and in-house programs support deep linking via a URL parameter or a link-builder tool. For SaaS affiliate programs, linking directly to a pricing page or feature-specific landing page can meaningfully lift EPC compared to routing all traffic through a homepage that requires additional navigation before users see what they came for.
Drip Campaign
A pre-written sequence of messages sent automatically on a schedule or in response to behaviour — a welcome series, an onboarding sequence, a re-engagement push. Because it runs once configured, a drip campaign delivers the right message at the right point in each person's lifecycle without anyone sending it manually. The most effective ones are triggered by what someone actually did or failed to do, rather than by elapsed time alone, which makes them feel timely instead of automated.
Dropshipping
A retail model where you never hold inventory: when a customer orders, the supplier ships the product directly to them, and you keep the difference between retail and wholesale. It removes upfront stock investment and warehousing, which is why it is a popular entry point into ecommerce. The trade-offs are severe and often underestimated — thin margins, no control over shipping times or packaging quality, and heavy competition because rivals can list the identical product from the identical supplier.
Dunning
The automated process of recovering failed subscription payments. When a card is declined — usually because it expired, hit a limit or was flagged — dunning retries the charge on a schedule and emails the customer to update their details. It matters more than most teams expect: involuntary churn from failed payments can account for a substantial share of total churn, and unlike voluntary churn it involves customers who still want the product. Smart retry timing and clear, non-alarming emails recover a large proportion of it.
E
EPC (Earnings Per Click)
EPC is an affiliate metric measuring average revenue generated per 100 clicks sent to an offer. Formula: total commissions ÷ total clicks × 100. It is a publisher-side metric for comparing offer profitability across different commission rates and conversion rates. EPC helps affiliates decide how much to pay for traffic. A 20% commission on a $200 product with a 1% conversion rate yields $0.40 EPC — the maximum CPC bid that breaks even. Networks often publish advertiser EPC to help publishers prioritize promotional focus. EPC varies significantly by traffic source quality; the same offer can show 2× difference in EPC between email and paid search traffic.
Edge Computing
Running code and serving data from locations physically close to the user rather than from one centralised region. Because the request travels a shorter distance, latency falls sharply — often the difference between a page feeling instant and feeling sluggish for users far from your origin. Edge platforms suit personalisation, authentication, redirects, A/B testing and API caching. The constraints are real: edge runtimes typically limit execution time, memory and available APIs compared with a traditional server.
Embeddings
Embeddings are numerical vector representations of text (or images, audio, or other data) that capture semantic meaning. Similar content produces vectors that are close together in high-dimensional space, enabling similarity search, clustering, and classification without keyword matching. Embeddings are the backbone of modern AI search, recommendation, and RAG systems. A query like "affordable residential proxies" embedded as a vector will surface semantically related results even if they share no exact words. Most embedding models output vectors of 768–3072 dimensions. Popular providers include OpenAI, Cohere, Google, and Cloudflare's built-in AI binding.
End-to-End Encryption
Encryption where data is encrypted on the sender's device and only decrypted on the recipient's, so nobody in between — including the service provider — can read it. This is stronger than encryption in transit, which protects data on the wire but leaves the provider able to read it on their servers. The trade-off is functionality: because the provider cannot see content, features like server-side search, previews and collaborative editing become harder, and losing your key can mean losing your data permanently.
Exit Node
The exit node is the final server or device in a proxy chain whose IP address the target website sees. In a residential proxy network, the exit node is the actual consumer device — a phone or desktop — routing your traffic through its connection to the internet. Exit node location determines the apparent geographic origin of your traffic. Understanding the full request path — your client → proxy infrastructure → exit node → target — matters for debugging latency issues: slow or overloaded exit nodes are often the bottleneck, not your bandwidth. Exit node health and speed vary significantly within large residential proxy pools.
Expansion MRR
Expansion MRR is the additional monthly recurring revenue generated from existing customers through upgrades, seat additions, add-on purchases, or moving to a higher pricing tier — without acquiring any new accounts. It is often the cheapest revenue to win because these customers already trust the product. Best-in-class SaaS teams aim for expansion MRR to outpace new-logo MRR within 2–3 years, at which point the business can grow even with flat new customer acquisition. Expansion MRR is tracked separately from new MRR to measure the health of the upsell and expansion motion.
F
Fine-Tuning
Fine-tuning is the process of continuing to train a pre-trained foundation model on a smaller, domain-specific dataset to specialize its behavior. Rather than training from scratch — which costs millions of dollars of compute — fine-tuning adapts an existing model for specific tasks at a fraction of the cost. Fine-tuning is used to make models follow specific response formats, adopt a particular tone, or develop expertise in a narrow domain such as medical documentation, legal drafting, or technical support. The risk: over-fine-tuning on narrow data can degrade general-purpose capabilities. RAG is often a better alternative when the goal is knowledge grounding rather than style or format specialization.
Foundation Model
A foundation model is a large AI model trained on broad, general data that can be adapted — via fine-tuning, prompting, or retrieval augmentation — for a wide range of downstream tasks. GPT-4, Claude, Llama, and Gemini are all foundation models. The "foundation" metaphor captures the idea that these models are starting points other applications build on, rather than training from scratch for every use case. Foundation model providers compete on benchmark scores, context length, safety, and price, while the application layer above them commoditizes rapidly as capabilities equalize.
Freemium
Freemium is a pricing model where a core set of features is permanently free, with additional functionality available on paid tiers. Unlike a free trial, freemium has no time limit — users can stay free indefinitely. The goal is to acquire a large user base that upsells over time. Freemium requires careful gate design: free must be valuable enough to attract real users, but constrained enough that growing teams need to upgrade. The median freemium-to-paid conversion rate is 2–5%; viral, team-based tools tend to convert higher because free users evangelize to colleagues who trigger team upgrades.
Fulfillment by Amazon (FBA)
A programme in which sellers ship inventory to Amazon's warehouses and Amazon handles storage, picking, packing, delivery, customer service and returns. FBA products become Prime-eligible, which materially improves conversion and Buy Box competitiveness. The costs are fulfilment fees per unit plus monthly storage, with sharply higher long-term storage fees for stock that does not sell — so FBA rewards accurate demand forecasting and punishes slow-moving inventory.
G
GDPR
The European Union's General Data Protection Regulation, which governs how personal data of people in the EU and EEA is collected, stored and used — regardless of where the company doing it is based. It requires a lawful basis for processing, meaningful consent for non-essential tracking, and grants individuals rights to access, correct, export and delete their data. Maximum fines reach the greater of €20 million or 4% of global annual turnover, which is why compliance is treated as a board-level risk rather than a legal footnote.
GRR (Gross Revenue Retention)
GRR measures the percentage of recurring revenue retained from existing customers — excluding any expansion. Where NRR can exceed 100% via upsell, GRR is capped at 100% because it only counts losses: churn and contraction. GRR is a purer signal of customer satisfaction than NRR. A company with high NRR but low GRR is papering over high churn with aggressive expansion. SaaS investors look at both together: GRR below 80% in enterprise or below 70% in SMB typically indicates a product or fit problem that upsell cannot fix.
Generative AI
Generative AI refers to AI systems that produce new content — text, images, audio, video, code — rather than just classifying or analyzing existing content. LLMs, image generators like Stable Diffusion and DALL·E, and audio synthesis models are all generative AI. The "generative" distinction matters for product design: these tools create artifacts that require human review and judgment, unlike a spam filter that returns a yes/no. Generative AI capabilities are expanding faster than governance norms, creating product opportunities and compliance risks for companies that deploy them in regulated industries.
Geo-Targeting (Proxy)
Geo-targeting in the proxy context means selecting IP addresses from a specific country, state/region, or city. Many web targets serve different content by geography — local prices, language versions, search results, streaming library availability, or ad creatives. Precise geo-targeting is critical for localized price monitoring, local SERP scraping, and ad verification. Country-level targeting is standard across all major providers; city-level targeting is available on premium plans and costs more due to smaller IP pools in specific locations. Some providers also offer ASN-level targeting to match traffic from specific ISPs.
Gross Margin
Revenue minus cost of goods sold, expressed as a percentage of revenue. For SaaS, COGS covers hosting, third-party APIs, payment processing and customer support — not sales and marketing. It is the single most telling indicator of business model quality, because it determines how much of each new dollar is available to fund growth. Software businesses typically target margins well above those of physical-goods businesses, and a persistently low SaaS gross margin usually signals hidden infrastructure or services costs.
Guardrails (AI)
Constraints placed around an AI system to keep its behaviour within acceptable bounds — input filters, output validation, refusal rules, permission scopes, spending caps and human approval gates for consequential actions. They exist because language models are probabilistic and will occasionally produce output that is wrong, unsafe or off-brand. Guardrails do not make a model correct; they limit the blast radius when it is wrong, which is what makes autonomous and agentic deployments viable in production.
H
Hallucination (AI)
Hallucination is when an LLM generates confident, fluent text that is factually incorrect — inventing citations, people, statistics, or events that do not exist. It is a fundamental limitation of how LLMs work: they generate plausible continuations of text, not verified truth. Hallucination is the primary risk in high-stakes LLM deployments — legal, medical, financial. Mitigation strategies include RAG (grounding answers in retrieved documents), structured output validation, tool-calling to query live data, and human-in-the-loop review. Hallucination rates vary significantly by model, task type, and how well the prompt constrains the output space.
Headless Browser
A headless browser is a real browser (Chromium, Firefox, WebKit) run without a visible UI, controlled programmatically via a driver like Playwright, Puppeteer, or Selenium. Headless browsers are used for automated testing, scraping JavaScript-rendered sites, PDF generation, and SSR smoke tests. They are slow and memory-hungry compared to raw HTTP, but necessary whenever a site loads data via fetch/XHR after initial HTML.
Headless CMS
A content management system that stores and serves content through an API without dictating how it is displayed. Removing the front end lets the same content feed a website, mobile app, kiosk and email system simultaneously, and lets developers use whatever framework they prefer. The trade-off is that non-technical editors lose WYSIWYG page building and live preview unless those are built separately, so headless suits multi-channel and developer-led teams more than small brochure sites.
I
ICP (Ideal Customer Profile)
An ICP is a data-driven description of the company type most likely to buy, retain, and expand — defined by industry, headcount, tech stack, geography, and buying behavior. Not to be confused with a buyer persona (which describes a person), an ICP describes the account. A precise ICP sharpens every growth lever: which channels to invest in, what to say in SDR outreach, which deals to qualify in or out, and where to focus product roadmap investment. Most early-stage SaaS companies under-invest in ICP definition and end up serving segments with incompatible needs — high CAC, high churn, and low expansion as a result.
IP Rotation
IP rotation is the practice of cycling through a pool of IP addresses during web requests to avoid detection and rate limiting. Rotation can happen on every request, on a time interval, or when a block is detected by monitoring response codes. Effective IP rotation strategies match the target site's behavior: aggressive rotation on large anonymous scrapes, sticky sessions where the site expects consistent IPs (logged-in accounts, shopping carts), and geographic targeting when content differs by region. Poor rotation — using too few IPs or too-predictable patterns — is one of the top causes of scrape failures at scale.
IP Whitelisting
IP whitelisting (or allowlisting) is a security control that grants access only to requests from pre-approved IP addresses. Proxy providers offer static IPs or username-based authentication so customers can add proxy IPs to their target's allowlist without exposing credentials in every request header. For web scrapers accessing APIs legitimately, whitelisting your proxy provider's egress IPs eliminates per-request auth overhead and reduces latency. Some enterprise SaaS products require IP whitelisting as part of security compliance — meaning automated workflows using them need stable, predictable IP ranges rather than rotating pools.
ISP Proxy
An ISP proxy (also called a static residential proxy) is an IP address assigned by an internet service provider to a data center — combining the speed of datacenter infrastructure with the legitimacy of a residential IP. These IPs appear residential to target sites but are hosted on fast hardware. ISP proxies sit between datacenter and true residential proxies on the speed-vs-detection spectrum: faster than mobile or rotating residential proxies, but less likely to be blocked than standard datacenter IPs. They are a good fit for tasks requiring both speed and low block rates — price monitoring, account creation, ticketing, and SERP data collection.
Inference (AI)
Inference is the process of running a trained AI model on new input to generate a prediction or output. When you send a message to Claude or generate an image with DALL·E, the serving infrastructure is performing inference — running the model weights against your input in real time. Inference cost and latency are the dominant operational concerns for production AI products. Larger models are more capable but slower and more expensive to serve. Techniques like quantization, batching, KV caching, and speculative decoding optimize inference throughput. Inference providers include Anthropic, OpenAI, Google, and specialized hardware clouds like Together.ai and Groq.
J
JavaScript Rendering
JavaScript rendering refers to executing a page's JavaScript before extracting data from it. Many modern sites load content via JS after the initial HTML response — the raw HTML contains only a scaffold that scripts populate with actual content after the page loads. Handling JS rendering requires a headless browser (Playwright, Puppeteer, or a cloud rendering service like Browserless or ScraperAPI). Rendering is the most significant cost multiplier in scraping: a headless browser uses 10–100× more CPU and memory than a raw HTTP request and is 5–20× slower. Identifying which pages require JS rendering vs. which can be scraped statically is a critical pipeline design decision.
K
Knowledge Cutoff
The date after which a model's training data ends. The model has no inherent knowledge of anything that happened later — new product releases, price changes, recent events — and if asked, may confidently produce outdated information rather than admitting the gap. This is why current-information tasks require retrieval: web search, RAG over your own documents, or tool calls that fetch live data. Always check the cutoff before relying on a model for anything time-sensitive.
L
LLM (Large Language Model)
An LLM is a deep learning model trained on vast amounts of text data to understand and generate human language. GPT-4, Claude, Gemini, and Llama are all LLMs. They work by predicting the next token in a sequence, building fluent, contextually appropriate text one step at a time. LLMs are the foundation of most modern AI products — from chatbots and writing assistants to code generation and data extraction. Their capabilities and limitations are shaped by training data, model size, and alignment techniques like RLHF. Understanding what an LLM can and cannot do reliably is the first step to building on top of one.
LTV (Customer Lifetime Value)
LTV (also called CLV or Customer Lifetime Value) is the total revenue — or more usefully, gross profit — a customer generates from sign-up to churn. The simplest formula: ARPU ÷ monthly churn rate. More precise models weight by cohort and account for expansion revenue. LTV drives every growth decision: how much you can bid for ads, whether an enterprise sales team makes economic sense, and how aggressively you can discount to land logos. SaaS companies with strong expansion motions often find LTV inflates significantly in months 12–24 as upsell kicks in.
Last-Click Attribution
An attribution model that assigns 100% of the credit for a conversion to the final touchpoint before it. It is simple, universally supported and unambiguous — which is why it remains the default in many tools — but it systematically overvalues bottom-of-funnel channels like branded search and coupon sites, while undervaluing the awareness activity that created the demand in the first place. Multi-touch models exist precisely to correct this distortion.
Latency (AI)
The delay between sending a request and receiving a response. For AI applications it is usually split into time-to-first-token, which governs how responsive the interface feels, and total generation time. Streaming output improves perceived latency dramatically because the user starts reading immediately rather than staring at a spinner. Latency generally trades against model capability and cost: larger models reason better but respond more slowly, which is why many products route simple requests to smaller, faster models.
Lead Magnet
Something valuable offered free in exchange for contact details — a template, checklist, calculator, report or mini-course. The best lead magnets solve a narrow, immediate problem for exactly the person you want as a customer, which both improves conversion and filters out unqualified signups. Generic giveaways attract volume but poor-quality leads. The strongest ones are naturally adjacent to your paid product, so consuming the free thing creates genuine appetite for the paid one.
Logo Churn
Churn measured by the number of customers lost, as distinct from revenue churn, which measures the money lost. The two can diverge dramatically and the gap is informative: losing many small accounts while retaining large ones produces high logo churn but modest revenue churn, whereas losing one enterprise account can barely move logo churn while devastating revenue. Reporting both is the only way to see what is actually happening in the base.
Lookalike Audience
An advertising audience built by an ad platform to resemble a source list you supply — typically your existing customers or highest-value users. The platform analyses shared characteristics and finds other users who match. Quality depends almost entirely on the seed list: a small list of your best customers usually outperforms a large list of all signups, because the model learns what a good customer looks like rather than what an average one looks like.
M
MRR (Monthly Recurring Revenue)
Monthly Recurring Revenue is the normalized monthly value of all active subscriptions. Annual plans are divided by 12. One-time fees, overages, and services are excluded. MRR is the canonical SaaS growth metric because it is predictable, comparable across pricing models, and directly drives runway and valuation math. Derived metrics include new MRR, expansion MRR, contraction MRR, churned MRR, and net new MRR.
Magic Number
A SaaS efficiency metric that measures how much new annual recurring revenue each dollar of sales and marketing generates, calculated by comparing the quarter-over-quarter increase in revenue against prior-period spend. A widely used rule of thumb treats results around 0.75 and above as a signal that growth spending is working and can be increased, while lower figures suggest fixing conversion, retention or targeting before pouring in more budget. Like all single metrics it is directional, not decisive.
Merchant of Record
The legal entity that sells to the customer, appears on their card statement, and carries responsibility for tax collection and remittance, chargebacks, refunds and compliance. Using a merchant-of-record provider means that entity takes on global sales tax, VAT and GST obligations on your behalf — a substantial burden for software sold internationally. The cost is a higher percentage fee than a plain payment gateway, and less direct control over the checkout and customer relationship.
Mixture of Experts (MoE)
A model architecture that contains many specialised sub-networks, or experts, but activates only a small subset for any given input. This decouples total parameter count from compute cost per token: the model can be very large in aggregate while remaining comparatively cheap and fast to run, because most of it stays dormant on each request. It is a major reason recent frontier models have improved capability without proportional increases in inference cost.
Mobile Proxy
A mobile proxy routes traffic through real mobile devices connected to carrier networks (3G/4G/5G), giving requests an IP address from a major telecoms provider like Verizon, T-Mobile, or Vodafone. Mobile IPs are the hardest to block because carriers assign them dynamically to millions of legitimate users. Mobile proxies command the highest prices in the proxy market — often 3–10× residential — because blocking a mobile carrier IP risks blocking real customers. They are essential for mobile-specific scraping, anti-fraud testing, ad verification on mobile platforms, and any target that actively distinguishes mobile from desktop traffic patterns.
Model Context Protocol (MCP)
An open standard for connecting AI models to external tools and data sources through a consistent interface. Rather than every application building bespoke integrations for every model, an MCP server exposes tools once and any MCP-compatible client — such as Claude, an IDE assistant or an automation platform — can use them. It matters because it turns integrations into reusable infrastructure, letting an assistant query your real systems with your existing permissions instead of relying on whatever it memorised in training.
Multi-Tenancy
Multi-tenancy is a SaaS architecture where a single instance of the software serves multiple customers (tenants), with data isolation between them. Most SaaS products are multi-tenant — one application layer handles all customers, separating their data in the database rather than running separate servers for each. The opposite is single-tenancy: each customer gets a dedicated instance. Multi-tenancy is far more cost-efficient to operate, but requires rigorous data isolation to prevent cross-tenant data leakage. Security-sensitive enterprise customers sometimes demand single-tenant or "private cloud" deployments, typically at a significant price premium.
Multi-Touch Attribution
An attribution approach that distributes credit for a conversion across several touchpoints in the customer journey rather than awarding it all to one. Models vary — linear splits credit evenly, time-decay weights recent interactions more heavily, position-based emphasises first and last. It produces a fairer picture of which channels genuinely contribute, but demands cleaner cross-channel tracking and is harder to explain to stakeholders than last-click.
Multimodal AI
AI that works across more than one type of input or output — text, images, audio, video — within a single model rather than chaining separate specialised systems. A multimodal model can read a screenshot and explain the error in it, or watch a clip and summarise what happened. Handling modalities natively usually preserves more context than converting everything to text first, which is why multimodal models tend to perform better on tasks where layout, tone or visual detail carries meaning.
N
NRR (Net Revenue Retention)
NRR measures how much revenue you keep and grow from existing customers over a period — after accounting for expansion (upsells, seat adds), contraction (downgrades), and churn. A 110% NRR means your existing customer base paid 10% more this period than last, even after all losses. NRR above 100% is the holy grail of SaaS: growth without acquiring a single new customer. It is the primary metric investors use to distinguish a leaky bucket from a compounding flywheel. Best-in-class B2B SaaS typically runs 115–130% NRR. Below 90% is a retention crisis.
Net Dollar Retention (NDR)
Revenue retained from existing customers over a period, including expansion, contraction and churn, but excluding new customers. Above 100% means the existing base grows by itself: upgrades and expansion more than replace everything lost. That is a powerful property, because it means revenue would still increase even if you acquired nobody new. It is among the metrics investors scrutinise most closely, since it captures product value and pricing power in a single number.
Net Promoter Score (NPS)
A loyalty metric derived from one question: how likely are you to recommend us, on a scale of zero to ten. Respondents scoring 9–10 are promoters, 7–8 passives and 0–6 detractors; the score is the percentage of promoters minus the percentage of detractors, producing a result between −100 and +100. Scores above 50 are commonly regarded as strong. Its real value is usually not the number but the free-text follow-up asking why, which is where the actionable detail lives.
O
OAuth
An open standard that lets you grant one application limited access to your account on another without ever sharing your password. When you click 'Sign in with Google', OAuth issues the app a scoped token instead of your credentials. Because tokens are scoped and revocable, you can grant read-only access to one app and revoke it later without changing your password or affecting anything else. It underpins nearly all modern third-party integrations.
P
PLG (Product-Led Growth)
PLG is a go-to-market strategy where the product itself is the primary driver of acquisition, activation, and expansion — rather than marketing campaigns or a sales team. Free trials, freemium tiers, and self-serve onboarding are the primary PLG mechanics. Users experience value before talking to anyone in sales. Successful PLG products generate leads from usage data: a team hitting 80% of their free quota is a warm sales signal without a single cold call. Slack, Notion, Figma, and Calendly are textbook examples. The trade-off: PLG requires the product to deliver value immediately, without hand-holding — which raises the bar for onboarding and activation.
Payment Gateway
The service that transmits card details from your checkout to the payment networks and returns an approval or decline, handling encryption and fraud screening along the way. It is distinct from a merchant of record: a gateway moves the money but leaves you legally responsible for tax, chargebacks and compliance. Gateways typically charge a percentage plus a fixed fee per transaction, and choosing one is largely a question of supported countries, currencies and payment methods.
Postback URL (S2S Tracking)
A postback URL (also called server-to-server or S2S tracking) is a server-side conversion notification: when a user converts, the advertiser's server pings the affiliate network's URL with conversion data. No cookies or pixels required — the data flows machine-to-machine. Postback tracking is more reliable than cookie-based tracking and immune to browser privacy restrictions. Most major affiliate networks support postback URLs for sophisticated publishers who need accurate server-side attribution — especially valuable for mobile apps, subscription billing where conversions happen weeks later, and cross-device purchase flows.
Product Qualified Lead (PQL)
A lead who has demonstrated buying intent through product usage rather than by filling in a form — someone who hit a plan limit, invited colleagues, or used a premium feature during a trial. PQLs typically convert at far higher rates than marketing-qualified leads because behaviour is stronger evidence than expressed interest. Acting on them requires instrumentation: the product must emit usage events that sales and lifecycle tooling can actually see and trigger on.
Prompt Engineering
Prompt engineering is the practice of crafting inputs to an LLM to reliably produce the desired output. It ranges from simple instruction phrasing to complex techniques like chain-of-thought prompting, few-shot examples, role-setting, and multi-step system prompt design. Good prompt engineering can dramatically change output quality without changing the model. In production AI systems, the system prompt is often the primary "code" that defines product behavior — making prompt engineering a core engineering discipline. Systematic prompt testing and version control are as important as any other part of the AI development workflow.
Prompt Injection
An attack in which malicious instructions hidden inside content the model reads — a web page, document, email or code comment — cause it to ignore its original instructions and do something else, such as leaking data or taking unauthorised actions. It is a fundamental issue for AI systems that process untrusted input, because models cannot reliably distinguish trusted instructions from text that merely looks like instructions. Mitigations focus on constraining privileges and validating actions rather than on filtering alone.
Proxy Authentication
How a proxy verifies you are entitled to use it. The two standard methods are username-and-password credentials sent with each request, and IP whitelisting, where the provider permits traffic only from IP addresses you register in advance. Credentials are more portable and work from changing networks, which suits cloud workers and laptops; whitelisting avoids embedding secrets in code but breaks the moment your own IP changes. Many providers support both simultaneously.
Proxy Pool
A proxy pool is the full inventory of IP addresses a proxy provider maintains for rotation and customer assignment. Pool size — often measured in millions of IPs for residential providers — determines how long a rotation strategy can run before repeating an IP against a given target. Not all pool IPs are equal: freshness (how recently tested), geographic distribution, and subnet diversity all affect scraping success rates. Premium proxy providers maintain "clean" pools by continuously testing IPs against common targets and retiring burned ones. Pool quality matters more than raw pool size for demanding scraping use cases.
Publisher & Advertiser
In affiliate marketing, the publisher is the affiliate — the website, newsletter, social account, or content creator driving traffic and earning commissions. The advertiser is the brand or company offering the product and paying commissions for conversions. Networks sit between the two, connecting publishers to advertiser offers and handling tracking and payments. Understanding this three-party structure clarifies who controls each lever: advertisers control offers, commission rates, creative assets, and cookie duration. Publishers control traffic quality, promotional placement, and audience targeting. Networks control tracking infrastructure, fraud protection, and payment terms. In-house affiliate programs cut out the network layer — increasing margin but requiring the advertiser to build and manage publisher relationships directly.
Q
Quantization (AI)
Reducing the numerical precision used to store a model's weights — for example from 16-bit to 8-bit or 4-bit — so it occupies far less memory and runs faster, often on cheaper hardware. The trade-off is a modest loss of accuracy that grows as precision drops. In practice moderate quantisation is close to lossless for many tasks, which is what makes it possible to run capable models on consumer GPUs and even laptops rather than datacentre accelerators.
Quick Ratio (SaaS)
A SaaS growth-efficiency metric comparing revenue gained to revenue lost: new plus expansion revenue divided by churned plus contracted revenue. It answers whether growth is genuinely compounding or merely offsetting leakage. A commonly cited benchmark treats a ratio around 4 as healthy — four dollars gained for every dollar lost. A low ratio alongside impressive top-line growth is a warning sign, because it means acquisition is masking a retention problem that will surface as growth slows.
R
RAG (Retrieval-Augmented Generation)
Retrieval-Augmented Generation combines a vector store (or hybrid BM25 + vector) with a language model. Relevant documents are retrieved at query time and injected into the model prompt, letting the LLM ground its output in specific source material instead of relying purely on pretraining. RAG is the standard architecture for enterprise knowledge-base bots, support automation, and internal copilots in 2026. Key failure modes: chunking strategy, embedding quality, and retrieval recall.
RLHF (Reinforcement Learning from Human Feedback)
RLHF is the training technique that aligns LLMs to human preferences — making them helpful, harmless, and honest. Human raters score model outputs, those scores train a "reward model," and the LLM is then optimized with reinforcement learning to generate outputs that score highly. RLHF is responsible for the conversational, instruction-following behavior of modern chatbots. Without it, a raw language model generates text that is statistically plausible but not reliably useful or safe. The quality of RLHF data — and the values encoded in rating guidelines — shapes how a model responds to edge cases, sensitive topics, and ambiguous requests.
Rate Limiting
Rate limiting is a server-side control that caps the number of requests an IP address, API key, or user can make within a time window. Exceeding the limit typically returns HTTP 429 (Too Many Requests) and may trigger a temporary or permanent block. For scrapers, working within or around rate limits is the practical difference between a sustainable pipeline and a blocked one. Well-designed scraping systems add exponential back-off, request jitter, and per-domain throttling to stay under detection thresholds. Many APIs publish their rate limits explicitly in documentation; websites rarely do, requiring empirical testing to find safe request rates.
Residential Proxy
A residential proxy routes traffic through an IP address assigned to a real home ISP subscriber, making requests look like organic consumer traffic rather than server traffic. Residential proxies are used where a target site blocks or rate-limits datacenter IPs — price monitoring, ad verification, sneaker bots, travel fare scraping, and SEO rank tracking. They cost 10-50x more than datacenter proxies because the IP inventory is sourced from SDKs, VPNs, and affiliate networks paid per GB of shared bandwidth. Key quality signals: IP pool size, geo targeting granularity, sticky session support, and ASN diversity.
Retargeting
Showing ads specifically to people who have already interacted with you — visited a page, viewed a product, abandoned a cart. Because these audiences have demonstrated intent, retargeting typically converts far better than cold prospecting, though it reaches a much smaller pool. The main risks are frequency fatigue from showing the same creative too often, and wasting budget on people who already converted, which sensible exclusion audiences prevent.
Return on Ad Spend (ROAS)
Revenue generated for every unit of currency spent on advertising, so a ROAS of 4 means $4 back for every $1 spent. It measures advertising efficiency in isolation, which is also its limitation: it ignores cost of goods, fulfilment and overhead, so a campaign can post an impressive ROAS while losing money overall. Profitable decision-making requires knowing your break-even ROAS given actual margins, and ideally judging campaigns on contribution profit rather than revenue.
Rotating Proxy
A rotating proxy automatically assigns a new IP address from a pool for each request — or at set time intervals — making large-scale web requests appear to come from many different sources. This prevents target sites from blocking a single IP address after detecting too many requests. Rotating proxies are fundamental to web scraping at scale. Most residential proxy providers enable rotation by default. The size and quality of the rotation pool — number of IPs, geographic spread, freshness — determines how long you can scrape a target before encountering blocks or CAPTCHAs. Some use cases require sticky sessions instead of pure rotation.
Rule of 40
The Rule of 40 states that a SaaS company's revenue growth rate and profit margin (usually EBITDA or free cash flow margin) should add up to at least 40%. A company growing at 60% YoY can burn at −20% margin. One growing at 20% needs to run at 20%+ margin. It is a quick heuristic investors use to balance growth and efficiency — particularly useful for comparing companies at different stages. As growth rates slow naturally with scale, the margin component must rise to compensate. Companies consistently above 40 are considered efficiently run; consistently below 30 triggers questions about unit economics.
Runway
How long a company can keep operating before it runs out of cash, calculated as cash on hand divided by net monthly burn. Twelve months of runway on $600,000 implies $50,000 net burn per month. It is the constraint behind nearly every startup decision, because fundraising takes months — so teams typically start raising with six-plus months remaining, and treat anything under six months as an urgent trigger to either cut burn or close funding.
S
SAML
An XML-based open standard that lets an identity provider pass authentication and authorisation data to a service provider, enabling single sign-on. In practice it is what allows an employee to sign in once with the corporate identity system and reach every connected application without separate passwords. It remains the dominant enterprise SSO protocol, which is why SaaS vendors almost always gate SAML support behind enterprise plans — it is frequently a hard procurement requirement.
SKU (Stock Keeping Unit)
A unique identifier assigned to each distinct sellable variant of a product. A shirt in three colours and four sizes is twelve SKUs, not one product, because each combination must be tracked, stocked and priced independently. Disciplined SKU structure underpins accurate inventory, purchasing and reporting; sloppy structure makes it impossible to answer basic questions like which variants actually sell. SKU count is also a common pricing dimension in ecommerce and inventory tools.
SLA (Service Level Agreement)
A contractual commitment defining the service level a provider will deliver — most often uptime, but also support response times and resolution targets — together with the remedy if they miss it. The remedy is usually a service credit rather than a refund of business losses, which is a crucial distinction: a 99.9% uptime SLA permits roughly 43 minutes of downtime per month, and the compensation for exceeding it rarely reflects what an outage actually costs you.
SOC 2
An auditing standard for how service organisations handle customer data, assessed against trust principles covering security, availability, processing integrity, confidentiality and privacy. Type I evaluates whether controls are properly designed at a point in time; Type II tests whether they actually operated effectively over a period, typically several months, and is therefore considerably more meaningful. For B2B software vendors it is frequently a prerequisite for closing enterprise deals.
SOCKS5 Proxy
SOCKS5 is a general-purpose proxy protocol that operates at the network layer, routing any type of TCP or UDP traffic — HTTP, HTTPS, FTP, or custom protocols. Unlike HTTP proxies, which only handle web traffic, SOCKS5 proxies work with any application that supports the SOCKS protocol. SOCKS5 also supports authentication and can tunnel UDP traffic, making it suitable for use cases beyond web scraping: gaming, VoIP, torrenting, and custom client applications. Most proxy providers offer both HTTP and SOCKS5 endpoints from the same IP pool, giving you flexibility to choose based on your client library or use case.
SSO (Single Sign-On)
A system letting users authenticate once and access multiple applications without signing in again to each. For employees it removes password fatigue; for IT it centralises control, so revoking one account instantly cuts access everywhere — which is why it is a standard security requirement for larger organisations. Common implementations use SAML or OAuth/OIDC. SaaS vendors typically place SSO on higher tiers, a practice sometimes criticised as the 'SSO tax'.
Sandbox Environment
An isolated copy of a system used for testing, where changes cannot affect real customers, live data or production billing. Sandboxes let teams trial integrations, rehearse migrations and train staff safely, and let developers exercise payment or API flows with test credentials rather than real money. The usual caveat is fidelity: sandboxes often lag production in features or data volume, so passing there does not guarantee identical behaviour live.
Scraping API
A managed service that handles the hard parts of web scraping — proxy rotation, browser rendering, CAPTCHA solving, retries and geographic targeting — behind a single API call. You send a URL and receive the page content or structured data, without maintaining your own proxy pool and headless browser infrastructure. It costs more per request than self-managed scraping but removes an entire category of maintenance, which usually wins for teams whose core product is not data collection.
Seat-Based Pricing
Seat-based pricing charges per user or per license — typically a flat monthly fee per active user. It is the most common SaaS pricing model for collaboration and productivity tools because it is easy to understand, predict, and align with HR headcount data. The challenge: enterprise buyers actively minimize seats, creating a ceiling on expansion MRR. Many SaaS companies layer usage limits or feature gates on top of seats to unlock additional revenue vectors beyond simple headcount growth. Seat-based models work best when every user gets consistent, measurable value from the tool.
Semantic Search
Search that matches on meaning rather than exact keywords, by converting both the query and the indexed content into vector embeddings and finding the nearest matches. A search for 'how do I stop paying' can surface a page about cancelling a subscription even with no shared words. It handles synonyms and phrasing variation far better than keyword search, but can miss exact-match needs like product codes — which is why production systems commonly blend both approaches.
Server-to-Server Tracking
Sending conversion data directly from your server to an ad platform's API rather than relying on a browser pixel. Because it does not depend on the user's browser, it survives ad blockers, cookie restrictions and tracking prevention that increasingly break client-side pixels, producing more complete and accurate attribution. It requires engineering work and careful deduplication when running alongside a pixel, but has become close to standard practice for serious advertisers.
Serverless
A model where you deploy functions and the platform handles all provisioning, scaling and server management, billing only for actual execution rather than for idle capacity. It scales from zero to high volume automatically, which suits spiky or unpredictable traffic and removes a large operational burden. The trade-offs are cold starts on infrequently used functions, execution time limits, and costs that can exceed dedicated servers under sustained heavy load.
Sticky Session
A sticky session (sometimes called a session-based proxy) holds the same exit IP for a configurable duration — typically 1, 10, or 30 minutes — so a sequence of requests appears to come from the same user. Sticky sessions are required whenever a target site tracks login state, cart state, or browser fingerprint across requests. Rotating proxies break these workflows because each request looks like a different user.
Sub-Affiliate
A sub-affiliate is an affiliate recruited by another affiliate (the master affiliate or super affiliate) who earns a cut of the sub-affiliate's commissions without doing the end-customer acquisition themselves. It creates a two-tier affiliate structure. Sub-affiliate programs are used to incentivize high-volume affiliates to recruit and mentor smaller publishers. The master affiliate typically earns 5–15% of sub-affiliate commissions. Networks offering sub-affiliate tiers attract super-affiliates who see recruiting as a scalable revenue stream alongside their own direct promotion activity.
Subscription Billing
Charging customers on a recurring schedule rather than per transaction, together with everything that entails: proration when plans change mid-cycle, upgrades and downgrades, trials, coupons, tax handling, invoicing, and dunning for failed payments. The edge cases are what make it hard — mid-cycle plan changes and refunds in particular — which is why most companies buy billing infrastructure rather than build it, and why billing bugs are a common source of customer trust damage.
Synthetic Data
Artificially generated data that mimics the statistical properties of real data without containing real records. It is used to train models where genuine data is scarce, expensive, sensitive or legally restricted, and to populate test environments without exposing customer information. The risk is fidelity: if the generated data misses patterns or edge cases present in reality, models trained on it perform worse than expected in production, so it is usually validated against real samples.
T
Temperature (AI)
A generation parameter controlling how much randomness a model uses when choosing each next token. Low values make output more deterministic and predictable, favouring the most probable continuation; higher values increase variety and surprise at the cost of consistency and factual reliability. Near-zero suits extraction, classification and code, where you want the same answer every time; higher settings suit brainstorming and creative writing, where repetition is the failure mode.
Time to Value (TTV)
How long it takes a new customer to experience the first real benefit of your product, measured from signup to that first meaningful outcome. It is one of the strongest predictors of retention, because motivation decays quickly after signup and every hour of setup before the first win increases the chance of abandonment. Shortening it usually means deferring configuration, shipping templates and sensible defaults, and importing existing data rather than asking people to start empty.
Token (AI)
In LLMs, a token is the basic unit of text the model processes — roughly ¾ of a word in English. "SaaSTweaks" might split into 3 tokens; "the" is usually 1. Token count determines both context window size and API pricing, since most LLM APIs charge per input token and output token separately. Efficient token usage — compressing context, truncating conversation history, caching repeated system prompts — is a practical cost-optimization skill for teams building LLM applications. Long documents, verbose prompts, or deep conversational history all consume tokens quickly. Prompt caching (supported by Anthropic and OpenAI) dramatically reduces costs for repeated system prompts.
Tokens per Second (TPS)
A throughput measure for language model inference — how many tokens the model generates each second. It determines how fast text appears and, combined with response length, how long a user waits. Roughly speaking, comfortable reading speed sits in the low tens of tokens per second, so anything above that feels immediate for chat, while batch and agentic workloads benefit from much higher throughput because they generate far more text without a human reading along.
Total Contract Value (TCV)
The full value of a contract across its entire term, including recurring subscription revenue plus any one-off fees such as implementation, training or professional services. A three-year deal at $50,000 per year with a $20,000 setup fee has a TCV of $170,000 and an ACV of $50,000. TCV shows the total commercial commitment, which matters for cash planning and commissions, while ACV is the right basis for comparing deals of differing lengths.
Tracking Pixel
A tracking pixel is a 1×1 transparent image embedded on a web page or in an email that fires a request to a tracking server when loaded — recording an impression, page view, or conversion event. The pixel URL carries user identifiers, session data, and event metadata as query parameters. Pixels are the traditional client-side tracking mechanism for affiliate conversions and ad campaign measurement. They are increasingly unreliable due to browser ad blockers, iOS privacy changes (ITP), and third-party cookie deprecation. Server-side postbacks and Meta's Conversion API (CAPI) are gradually replacing pixels for conversion attribution in programs that need high accuracy.
Two-Factor Authentication (2FA)
A security control requiring a second proof of identity in addition to a password — typically a time-based code from an authenticator app, a hardware key, or a push notification. Because it defeats attacks that rely on a stolen or reused password alone, it blocks the overwhelming majority of account takeovers. Not all factors are equal: hardware keys and authenticator apps resist phishing far better than SMS codes, which are vulnerable to SIM-swap attacks.
U
UTM Parameters
Tags appended to a URL that tell analytics tools where a visitor came from — source, medium, campaign, and optionally term and content. They are how you distinguish a click from a newsletter, a paid ad and a partner link that all land on the same page. Their weakness is discipline: inconsistent capitalisation or naming fragments the same campaign across multiple rows in reporting, so teams that rely on UTMs almost always maintain a documented naming convention.
Uptime
The percentage of time a service is available and functioning, usually measured monthly and often committed to in an SLA. The practical meaning of each additional nine is easy to underestimate: 99.9% permits roughly 43 minutes of downtime per month, while 99.99% permits about four. Read the definition carefully too, since providers frequently exclude scheduled maintenance, third-party failures and force majeure from the calculation.
Usage-Based Pricing (UBP)
Usage-based pricing charges customers based on what they consume — API calls, data volume, active seats, or compute hours — rather than a flat monthly fee. Also called consumption-based or metered pricing. It lowers the entry barrier because customers start small and pay more as they get more value. UBP aligns vendor and customer incentives cleanly: as customers succeed, revenue grows. Companies like Stripe, Twilio, and Snowflake use UBP as a core growth engine because customers naturally expand as their business scales. The trade-off is revenue volatility — usage can spike or drop significantly month to month.
User Agent
The user agent (UA) is a string sent in HTTP request headers that identifies the client software — browser type, version, and operating system. Websites use it to serve the right content format and as a basic bot-detection signal. Scrapers that send default HTTP library user agents (like "python-requests/2.28") are trivially detectable. Rotating realistic UA strings matching current browser versions and their exact header ordering is a baseline anti-bot countermeasure. Full browser fingerprinting goes beyond UA to include screen resolution, fonts, WebGL renderer, timing data, and hundreds of other signals — making a convincing browser impersonation significantly harder than swapping a header string.
V
Vector Database
A vector database stores and indexes high-dimensional embedding vectors, enabling fast similarity search across millions of records. Unlike a traditional database optimized for exact matches, a vector DB retrieves the nearest semantic neighbors to a query vector — finding the most relevant documents even when they share no keywords with the query. Vector databases are the retrieval layer in most RAG systems. When a user asks a question, their query is embedded and matched against pre-indexed document embeddings, retrieving the most relevant chunks to include in the LLM prompt. Popular options include Pinecone, Weaviate, Qdrant, pgvector (Postgres extension), and Cloudflare Vectorize.
Vector Embedding
A numerical representation of content — text, an image, audio — as a list of numbers positioned in a high-dimensional space such that similar meanings sit close together. Because similarity becomes a distance calculation, embeddings power semantic search, recommendations, clustering and retrieval-augmented generation. Quality depends on the embedding model and on how content is chunked before encoding: chunks that are too large blur meaning, while chunks that are too small lose context.
W
Web Crawling
Web crawling is the automated traversal of websites by following links to discover and index content. A crawler starts from a seed URL, parses all links on that page, follows them, and repeats — building a map of an entire site or web graph. Search engine bots like Googlebot are web crawlers. Crawling is distinct from scraping: a crawler discovers URLs, a scraper extracts data from them. For competitive intelligence or market research, crawling is typically the first phase — discovering all relevant URLs — before targeted scraping extracts specific data from each. Robots.txt files signal which paths crawlers should skip.
Web Scraping
Web scraping is the automated extraction of structured data from websites — product prices, review text, contact information, job listings, or any publicly visible content. A scraper issues HTTP requests, parses the returned HTML or JSON, and extracts target fields into a usable format. Modern web scraping is an arms race: websites deploy bot detection, CAPTCHAs, JavaScript challenges, and dynamic rendering to block automated access, while scrapers counter with proxies, browser automation, and behavioral mimicry. The legal landscape varies by jurisdiction and target site — always check a site's terms of service and applicable law before scraping at scale.
Webhook
An automated HTTP request a service sends to a URL you specify when a particular event occurs, so you receive updates the moment they happen instead of repeatedly polling an API to ask. It is the difference between being told and having to check. Because delivery happens over the public internet, robust receivers verify the request signature, respond quickly and process asynchronously, and handle duplicates — since most providers retry and may deliver the same event more than once.
Webhook Retry
The mechanism by which a provider re-sends a webhook that failed to deliver — because your endpoint was down, timed out or returned an error — usually on an exponential backoff schedule over minutes or hours before giving up. It protects against transient failures, but has an important consequence: your endpoint must be idempotent, meaning the same event processed twice produces the same result. Otherwise retries cause duplicate orders, double charges or repeated emails.
White-Labeling
White-labeling is when a vendor allows customers or partners to rebrand the product as their own — replacing logos, colors, and domain with the buyer's brand identity while the underlying technology stays the vendor's. The end customer sees the partner's brand, not the original vendor. Common in SaaS platforms, white-labeling is a B2B2C distribution model: an agency or reseller packages your tool under their brand for their customers. It accelerates distribution without building partner marketing, but creates support complexity and can commoditize the core product if the vendor becomes invisible to end users.
Z
Zero Trust
A security model that assumes no user, device or network location is inherently trustworthy, so every request is authenticated and authorised regardless of origin. It replaces the older castle-and-moat approach, where anything inside the corporate network was implicitly trusted — an assumption that fails badly with remote work, cloud services and contractors. In practice it combines strong identity, device posture checks, least-privilege access and continuous verification rather than a single perimeter check.
Zero-Shot vs. Few-Shot Learning
Zero-shot means prompting an LLM to perform a task without providing any examples — relying entirely on the model's training. Few-shot means providing 2–5 input/output examples in the prompt before the actual task, giving the model a format to follow. Zero-shot is simpler and uses fewer tokens; few-shot often produces more consistent output on structured tasks like classification, extraction, or formatting. The choice between them is a practical engineering decision: test both with your specific task and model, measure output quality, and choose accordingly. Many production systems combine both — a zero-shot instruction with a few worked examples embedded in the system prompt.