The terms this product is built on, defined precisely enough to cite. Contract words mean exactly what the API enforces; category words get the working definition we measure against.
Start with the contract words like Envelope and absence-is-data, or the category words like GEO. Every definition is enforced somewhere in the Envelope.
The contract rule that an AI surface returning nothing for a query is a valid, reportable result, not an error, and never a fabricated answer. If no AI Overview appears, the capture still completes: the surface is recorded as absent and the job is flagged, so downstream metrics can count the absence. Coercing absence to zero, retrying it away, or inventing content would corrupt visibility data.
Full entry: how it works in the API →also: AI bot · answer-engine crawler
The automated agents that answer engines and AI assistants use to fetch, render and index pages before citing them: GPTBot and OAI-SearchBot for ChatGPT, PerplexityBot for Perplexity, ClaudeBot for Claude, Google-Extended for Google’s AI Overviews and AI Mode. Each engine runs its own crawler under its own user agent, and a page that disallows one in robots.txt has opted out of citation by that engine specifically, whatever else the page gets right stops mattering to it.
Full entry: how it works in the API →also: Google AI Mode
Google’s fully conversational search experience, a step past AI Overviews, with its own deeper retrieval, follow-up questions and multi-query synthesis, replacing the results page with a composed answer for the queries it handles. It runs its own retrieval, so what it cites and how it phrases an answer can differ from both classic Search and AI Overviews for the identical query.
Full entry: how it works in the API →also: Google AI Overview · AIO
Google’s AI-generated summary block shown above the classic results for many queries, composed by Google’s models from retrieved pages, with source links attached. Because it sits at position zero and answers the question directly, it reshapes what gets read and clicked, and there is no official Google API for it, so capturing one means rendering the real results page and extracting the block.
Full entry: how it works in the API →also: share of model · answer share
A brand’s proportion of presence in AI-generated answers relative to competitors: across a set of captured answers, how often each brand is mentioned or cited, in what position, and with what sentiment. Because AI answers differ by surface, phrasing and region, share of voice is only meaningful when computed from the actual rendered answers, captured and compared over time.
Full entry: how it works in the API →How often, how prominently, and in what light a brand, product or site appears in AI-generated answers across surfaces and regions. It is the AI-era counterpart of search visibility: instead of rank positions, the units are mentions inside answers, citations of your pages, position within the answer, and sentiment, all of which vary by surface, query phrasing and geography.
Full entry: how it works in the API →One typed unit of a normalized AI answer, a paragraph, heading, list, code block or quote. Answers arrive as prose from the surface, but block structure exists so a downstream pipeline can work at the level of a specific paragraph or list item, without re-parsing markdown or guessing at sentence boundaries.
Full entry: how it works in the API →also: AEO
Optimizing content to be selected as the answer by answer engines, systems that respond with one composed answer instead of a list of links: Perplexity, Google’s AI Overviews, ChatGPT with search. AEO overlaps heavily with generative engine optimization; AEO emphasizes being the extracted, cited answer, while GEO names the broader practice across generative surfaces. In use, the two terms are often interchangeable.
Full entry: how it works in the API →also: search batch
Submitting many prompts, across surfaces and regions, in a single request instead of one call per query. A batch fans out into independent child captures, each billed and completed on its own, so a nightly sweep of hundreds of prompts is one request cycle rather than hundreds of round trips.
Full entry: how it works in the API →Watching what AI surfaces say about a brand for content that is false, defamatory, or exposes the company to regulatory or reputational risk, a regulated claim attributed incorrectly, a competitor’s product recommended in a brand’s place, an unsafe or non-compliant instruction tied to a product name. It works as a discipline only when each observation is a structured, repeatable capture you can diff over time, not a one-off screenshot nobody can re-derive.
Full entry: how it works in the API →also: consumer-UI capture
Rendering the actual consumer-facing app or page in a real browser session and reading the result, instead of calling a vendor’s sanitized model API and presenting that as equivalent. The two are not interchangeable: a model API answer skips whatever ranking, presentation, ads and UI-level behavior the real product applies, so it is not what a real user of that surface actually sees.
Full entry: how it works in the API →The billing rule that a capture only spends credits when it actually completes with a result, a failed acquisition or a driver outage costs nothing. It matters because a metered API without this rule punishes exactly the failures that are least the customer’s fault, and it makes fail-loud behavior (an honest error instead of a fabricated mock) commercially safe to enforce.
Full entry: how it works in the API →also: citation
A source an AI answer actually used, as distinct from one it merely fetched and discarded. The distinction is the difference between "the engine saw this page" and "the engine’s answer depends on this page", and only the second is what GEO and citation-tracking work needs to measure, because only the second is what a reader is actually being told to trust.
Full entry: how it works in the API →also: API credit
The unit every capture is billed in, at a cost that varies per surface rather than one flat per-call price, a harder-to-reach or more compute-intensive surface can cost more than a simple one. Credits are the whole billing model: no per-field surcharges, no separate charges for artifacts, one number per successful capture.
Full entry: how it works in the API →The single, versioned response contract returned for every capture, regardless of surface. An Envelope carries the job header (id, surface, region, status, warnings), provenance (how each claim about the capture is known), the answer (text, markdown, blocks) and the structured evidence extracted from it (sources, fan-out). On consumer-UI (browser/scrape) surfaces you can additionally opt in to a proof-of-page HTML snapshot with include.html, surfaced as a top-level html URL. One shape to parse, diff and store, instead of N per-provider formats.
Full entry: how it works in the API →also: query fan-out
The set of background search queries an AI surface issues to answer a single prompt. Surfaces like ChatGPT with search and Google’s AI Mode decompose one question into several sub-queries before composing the answer, and some of them print those sub-queries right in the answer. Where a surface exposes its fan-out this way, the queries land in the captured answer text, revealing which searches your content must win to be retrieved and cited.
Full entry: how it works in the API →also: GEO · LLMO
The practice of improving how generative AI systems, ChatGPT, Perplexity, Google’s AI Overviews and AI Mode, describe, cite and recommend your content or brand. Where classic SEO optimizes for ranked links, GEO optimizes for being used in the composed answer: cited as a source, named in a recommendation, and matched to the fan-out queries the engine actually runs underneath.
Full entry: how it works in the API →also: regional capture
Running the identical query from a specific country (and optionally a city) so the answer reflects what that market actually sees, AI answers, AI Overviews and SERPs vary by region, sometimes sharply, and a single-market capture cannot reveal that. Regions are a request parameter, not a separate product: the same Envelope shape returns, stamped with which market produced it.
Full entry: how it works in the API →also: logged-out capture
A browser capture performed without signing in to the surface being captured, the default and, for most live surfaces today, the only mode. It is honestly labeled because a logged-in session can personalize an answer in ways a guest session never sees, and conflating the two would corrupt any comparison across captures.
Full entry: how it works in the API →also: AI hallucination
An AI system generating an answer, citation or fact that is fabricated rather than grounded in anything real, the single biggest reason technical buyers distrust AI-generated content. A monitoring tool built on top of AI answers inherits this risk twice over: once from the surface being captured, and again if the capture pipeline itself ever fills a gap with something plausible instead of admitting it saw nothing.
Full entry: how it works in the API →A client-supplied identifier attached to a request so that submitting it again, after a timeout, a retry, or a cron job firing twice, resolves to the original result instead of creating a duplicate capture or double-billing. It is the mechanism that makes a scheduled, unattended capture pipeline safe to retry blindly.
Full entry: how it works in the API →also: structured data · schema markup
A JSON-formatted way of embedding schema.org structured data in a page so search engines and AI systems can read the page’s meaning, its type, its facts, its relationships to other pages, without inferring it from prose. It is a machine-readable annotation layered on top of the human-readable page, not a replacement for it.
Full entry: how it works in the API →also: LLM manifest
A plain-text file at a site’s root, written for AI systems and agents rather than human browsers: a short summary of what the product does, the honest facts worth stating up front, and links to the pages that matter most. It is the AI-era counterpart to robots.txt, not a directive, a briefing.
Full entry: how it works in the API →An AI surface that requires a signed-in account before it will answer: the Claude consumer app, Meta AI, DeepSeek, Amazon Rufus, and Grok. Anonymous proxy scraping structurally cannot reach them. There is no session to borrow. Reaching a login-walled surface takes either the vendor’s official API (a surrogate for the UI) or a real authenticated browser session.
Full entry: how it works in the API →also: observed vs inferred
The record of how a capture was acquired and how each claim about it is known. Observed facts are read directly from the surface; inferred ones are estimates and must be flagged as such with a confidence score, a UI model label is never silently equated with an official model ID. Provenance covers login state, region requested versus effective, and whether the surface was present at all.
Full entry: how it works in the API →also: SEO rank tracking · position tracking
Recording where a page ranks among the ten blue links for a given query, over time, the classic SEO monitoring discipline. It answers "where do I place on the results page" and nothing about what an AI-composed answer says, which is a structurally different question once an AI Overview or a chat answer sits above or instead of the ranked list.
Full entry: how it works in the API →also: schemaVersion · contract versioning
A stamp on every capture recording exactly which version of the Envelope contract, capture driver, answer normalizer and provenance layer produced it, so a stored Envelope from six months ago can be understood correctly even after the contract itself has evolved. Without it, a schema change silently reinterprets old data instead of leaving it honestly attributable to the version that wrote it.
Full entry: how it works in the API →also: mention sentiment
How positively, neutrally or negatively an AI answer characterizes a named brand, a qualitative read that matters as much as whether the brand was mentioned at all, since being named in a comparison as the worse option is a different outcome than being recommended. It is derived from the answer’s actual phrasing around the mention, not assumed from mention count alone.
Full entry: how it works in the API →also: search engine results page
The page a search engine actually renders for a query, organic results, and increasingly, AI-generated modules sitting above or inside them. Once AI Overviews and AI Mode became part of the page, "the SERP" stopped meaning just ranked links; it is now the whole rendered surface an answer engine or organic result competes for.
Full entry: how it works in the API →also: sync mode · synchronous capture
Requesting a capture and blocking for the finished Envelope inline in the same response, instead of receiving a job id to poll, built for the moment a pipeline step has a user actually waiting on a result. It is the same endpoint, the same body and the same Envelope as the default async path; only the response timing changes.
Full entry: how it works in the API →also: HMAC-signed webhook
An HTTP callback the API sends to a URL you register the moment a capture finishes, instead of you polling for it. For a monitoring pipeline running many prompts across many surfaces and regions, webhooks turn "was this done yet" from a poll loop into an event your system reacts to as it happens.
Full entry: how it works in the API →also: no-click search
A search where the user’s question gets answered directly on the results page or inside a chat answer, with no click to any underlying website required to satisfy it. AI Overviews, AI Mode and answer-engine chat responses are built to produce exactly this outcome, which is precisely why being the cited source inside the answer matters more than ranking below it.
Full entry: how it works in the API →Every definition above is enforced somewhere in the Envelope. See the surfaces you can capture, or run a query and read the fields yourself.