AI answer tracking: consumer interfaces vs model APIs

How to decide in two questions

First question: what should the number mean? If it should describe what a person sees when they ask ChatGPT, the Gemini app, Google AI Mode, AI Overviews or the Claude app, you are measuring a consumer surface. If it should describe what a product returns when it calls a provider's search tool, you are measuring an API route. The sources checked on 8 November 2026 describe the app features and the API tools in separate documents with separate controls (OpenAI API guide, ChatGPT help, Gemini API guide, AI Mode help, Claude API docs, Claude app help). None of them says the two return the same answer, and this research ran no queries, so it measured no difference either. This page claims neither a match nor a gap.

Second question: what will you do with the results? Showing, storing and reusing grounded API output is governed by each provider's own terms, and the three differ. The section on access, display and reuse quotes them provider by provider. To choose a monitoring product, see Otterly vs Peec vs Profound and Ahrefs Brand Radar vs Semrush. This page is about how to label and read what you collect.

What each surface is

OpenAI documents two routes. For developers, the Responses API offers a web_search tool, and the current guide recommends gpt-6-astra for new integrations (OpenAI web search guide). For people, ChatGPT search is available on ChatGPT Free, Go, Plus, Pro, Business, Enterprise and Edu (OpenAI help). The sources checked do not establish which backend serves the consumer feature.

Google has more surfaces. The Gemini app, AI Overviews and AI Mode are separate, and Google says AI Mode and AI Overviews "may use different models and techniques" (AI features and your website), so I do not treat them as one model. AI Mode gives an AI-powered response with links to the web (AI Mode help). For developers, the Gemini API has a google_search tool, and the Interactions example in the guide uses gemini-3.8-flash (Grounding guide). That is a documentation example, not a statement about what the Gemini app runs. Two response formats appear in the docs: Interactions returns inline annotations with search suggestions (Interactions), while generateContent returns structured citation data (generateContent). Keep the two apart when you parse them.

Anthropic's app web search works with a list of named models (Claude app help). On the API, web search is a versioned tool: web_search_20250305 is basic web search, and the documentation lists later versions (Claude API docs). The app route offers model choices and the API route selects a model and a tool version. The sources checked do not say the API call equals the full app.

How these facts were checked

On 8 November 2026 an AI-assisted research browser read each provider's public documentation, help pages, pricing pages and terms. No queries were run, no API calls were made, no accounts were created, nobody signed in and no money was spent. Documentation examples are examples, not executed tests. So I can tell you what the providers document, and I cannot tell you what any surface returned for any prompt.

The terms are the public versions the browser landed on. OpenAI's individual terms observed are the rest-of-world version, effective 1 January 2026; the EEA, UK and Switzerland have separate terms (OpenAI Terms of Use). The Google Terms observed are the Pakistan version, effective 30 July 2026 (Google Terms of Service), and the Gemini API Additional Terms are effective 23 March 2026 (Gemini API terms). This is not legal advice, and an agreement negotiated with a provider can differ. Reading a provider's terms was research. It is not evidence that anyone has the right to run a tracking service.

Citations on each surface

Cited answers are documented on several surfaces here. A citation format says nothing about whether the answers or the cited pages match across surfaces.

Location and personalization

In my view, a consumer observation with no location and no account state attached is hard to interpret, because the providers document several inputs that can shape what a person sees. That is an editorial reading of the documentation, not a tested result.

Search tool pricing as listed on 2026-11-08

The units differ: calls, requests and searches. On Google one prompt can use more than one billable query (Gemini API pricing). I did not build a cost per answer, because the pages give none and this review ran no queries. Consumer plan prices were not checked for any provider, and they are not API price units.

Access, display and reuse terms, by provider

These quotes are exact, with spacing normalised for reading. Read each provider on its own. Nothing here transfers from one provider's terms to another's.

OpenAI

Consumer: the individual terms list as a prohibited use "Automatically or programmatically extract data or Output (defined below)" (OpenAI Terms of Use). Third Party Services and Third Party Output "are subject to their own terms" (OpenAI Terms of Use).

API: "When your interface displays web results or information from them, you must show inline citations as prominent, clickable links" (OpenAI web search guide). The Services Agreement says: "As between Customer and OpenAI, to the extent permitted by applicable law, Customer: (a) retains all ownership rights in Input; and (b) owns all Output" (Services Agreement). It also bars use of the services or customer content in a way that violates third parties' rights (Services Agreement), and by accessing a Third-Party Service the customer agrees to its terms (Services Agreement). A grounding-specific fixed storage period and a blanket redistribution ban were not found in the web search guide, the Services Agreement and the Service Terms checked on 8 November 2026. That describes those documents. It is not permission to store or redistribute.

Google

Consumer: Google's Terms prohibit "using automated means to access content from any of our services in violation of the machine-readable instructions on our web pages (for example, robots.txt files that disallow crawling, training, or other activities)" (Google Terms of Service). The prohibition is conditioned on those instructions. It is not a blanket ban on automation, and it does not show that a given method is authorized.

API, under the Gemini API Additional Terms: you "will only use Grounding with Google Search in an application that is owned and operated by you and will only display the Grounded Results with the associated Search Suggestion(s) to the end user who submitted the prompt" (Gemini API terms). Unless Google permits it in writing, you will not modify the Grounded Results or Search Suggestions or intersperse other content with them, and will not place interstitial content between a Link or Search Suggestion and its destination page (Gemini API terms). You also will not "cache, frame, syndicate, resell, analyze, train on, or otherwise learn from Grounded Results or Search Suggestions" (Gemini API terms). Using grounding to extract or collect Grounded Results, Search Suggestions or Links for another purpose is a violation, for example "using programmatic or automated means to collect Links" (Gemini API terms).

Storage is a prohibition with listed exceptions: you will not "copy, store, or implement any click tracking, Link-tracking or other monitoring of Grounded Results or Search Suggestions, except that" the exceptions apply (Gemini API terms). One exception lets you "copy and store, for up to two (2) years, the text of the Grounded Result(s)" that you displayed only to evaluate and optimize the display, or that sit in an end user's chat history only so that user can view it (Gemini API terms). The full clause also allows temporary function-call refinement under deletion and final-display conditions (Gemini API terms). End users may copy and store individual displayed results if you do not allow them to be accessed or collected by automated or programmatic means or used to create a database (Gemini API terms). Results and Suggestions may also be stored for the minimum time necessary to comply with law (Gemini API terms). So the terms are neither a blanket ban on storage nor a general two-year allowance. These are the Gemini Developer API terms themselves; I did not import any detail from Vertex AI documentation.

Anthropic

Consumer: the Consumer Terms list this prohibited use: "Except when you are accessing our Services via an Anthropic API Key or where we otherwise explicitly permit it, to access the Services through automated or non-human means, whether through a bot, script, or otherwise" (Consumer Terms).

API: the Commercial Terms give permission to use the Services "including to power products and services Customer makes available to its own customers and end users", subject to the terms (Commercial Terms). The customer may not "access the Services to build a competing product or service, including to train competing AI models or resell the Services except as expressly approved by Anthropic" (Commercial Terms). As between the parties and to the extent permitted by law, the customer owns its Outputs (Commercial Terms). The customer is responsible for evaluating whether Outputs are appropriate for its use case, including where human review is appropriate, before using or sharing them (Commercial Terms). A grounding-specific fixed storage period and a blanket redistribution ban were not found in the Commercial Terms, the Service Specific Terms and the web search documentation checked on 8 November 2026. That is not clearance. Source and publisher rights, and applicable law, are separate questions.

What the sources do not establish

• OpenAI: the sources checked do not say that Responses API web_search output matches what ChatGPT search shows (API guide, ChatGPT help).
• Google: the sources checked do not say that Gemini API grounding reproduces the Gemini app, AI Overviews or AI Mode (Grounding guide). Google itself says AI Mode and AI Overviews may use different models and techniques (AI features and your website). A grounded API response should not be labelled as an observed AI Overviews or AI Mode answer.
• Anthropic: the app lists several model choices (Claude app help) and the API route selects a model and a tool version (Claude API docs). The sources checked do not say the two are the same experience.

"Not found in the sources checked on 8 November 2026" is a coverage note about those pages. It is not evidence that the answers differ, and it is not evidence that they match. This research measured neither.

A record that keeps the label honest (my editorial view)

This is my suggestion, not a provider standard, and I have not tested whether it changes any result. For every observation you keep, record the surface (for example ChatGPT search, the Gemini app, AI Mode or the Claude app, or the API route and tool name), the timestamp, the model and tool version where you can select them, the geography, the account and memory state, and the terms you relied on to collect and keep it. The reasons come from documentation: location and memory inputs differ by consumer surface (ChatGPT help, ChatGPT help, Gemini Apps Privacy Hub, AI Mode help, Claude privacy page, Claude memory help), the API tool versions differ (Claude API docs), and Google's terms set a purpose for each kind of storage (Gemini API terms).

When you report, say which kind of record it is: a consumer observation on a named surface, or an API result from a named tool and version. Keep them in separate columns.

What each route does well and badly

Consumer observations do well at describing the surface people use. Claude's app help says every web-search response includes citations, and ChatGPT help says responses that use web search may include citations (Claude app help, ChatGPT help). AI Mode help documents links to the web, not a citation guarantee (AI Mode help). Their limits: the documented inputs include location, memory and account settings (ChatGPT help, Gemini Apps Privacy Hub, AI Mode help, Claude memory help); sources are not promised on every response (Gemini Apps help, ChatGPT help); automated collection is restricted by the consumer terms of OpenAI and Anthropic, and conditioned on machine-readable instructions in Google's (OpenAI Terms of Use, Consumer Terms, Google Terms of Service); and the exact backend is not established for ChatGPT search (ChatGPT help).

API results do well at control and pricing clarity. The docs describe location fields (OpenAI guide, Claude API docs), citations (OpenAI guide, Claude API docs, Gemini grounding guide), versioned tools (Claude API docs) and a listed price per search unit (OpenAI pricing, Gemini API pricing, Claude API docs). Their limits: display and reuse conditions (OpenAI guide, Gemini API terms, Gemini API terms, Gemini API terms, Commercial Terms); on Google, one prompt can be billed as several queries (Gemini API pricing); no source checked says the output matches the consumer surface; and a location setting was not found in the Gemini grounding guide checked, which is a coverage note and not an absence claim.

Who should pick which

• Pick consumer observations if the question is what a person in a given place and account state could see, and you can show your collection method fits the access clauses above (OpenAI, Google, Anthropic). Label every record as a consumer observation.
• Pick API results if you want parameters you control, such as location fields and tool versions, and a priced unit. Label them as API results with the tool and version (OpenAI, Google, Anthropic). Do not describe them as what ChatGPT, Gemini, AI Mode, AI Overviews or Claude users saw.
• If you plan to store, show or sell results, read the provider's reuse terms first. Google's grounding terms list narrow storage purposes (Gemini API terms, Gemini API terms). The OpenAI and Anthropic documents checked contain no grounding-specific storage period, which is not clearance.
• If you need both, report them side by side and do not blend them into one figure. That is my view, not a tested method.

FAQ

Does a Responses API result show the answer a ChatGPT search user saw?

The sources checked do not say it does. OpenAI documents the web_search tool in the Responses API (OpenAI guide) and ChatGPT search in a separate help article (ChatGPT help). This research ran no queries, so it measured no difference either.

Does Gemini API grounding reproduce AI Overviews or AI Mode?

The sources checked do not say it does. Google documents the google_search tool for the API (Grounding guide) and says AI Mode and AI Overviews may use different models and techniques (AI features and your website). That is not an API-versus-app comparison.

Must Search Suggestions be displayed with Gemini grounded answers?

Under the Gemini API Additional Terms effective 23 March 2026, you display the Grounded Results with the associated Search Suggestion(s), only to the end user who submitted the prompt, in an application you own and operate (Gemini API terms). You may not modify them or intersperse other content unless Google permits it in writing (Gemini API terms).

Can grounded answers be stored in an AI visibility database?

It depends on the provider. Google prohibits copying and storing Grounded Results except for listed purposes (Gemini API terms, Gemini API terms), and end-user copies must not be used to create a database (Gemini API terms). For OpenAI and Anthropic, no grounding-specific storage period was found in the documents checked on 8 November 2026, which is not permission; third-party rights and output-use duties still apply (Services Agreement, Commercial Terms). This is not legal advice.

Why can one prompt create more than one search charge?

On Google, a request to Gemini may result in one or more queries to Google Search, and each individual query performed is charged (Gemini API pricing). Prices are as listed on 2026-11-08: $14 per 1,000 requests after 5,000 free search requests per month, shared across Gemini 3 and newer models on the paid tier, subject to a daily cap of 1,500 requests per day (Gemini API pricing).

How should a report label API results versus consumer observations?

In my view, name the surface or API route, the tool and model version where known, the date, the location and the account state, and keep the two kinds of record in separate columns. The providers document location, memory and tool-version inputs (ChatGPT help, Gemini Apps Privacy Hub, Claude API docs). I have not tested this labelling.

Disclosure, gaps and changelog

No account testing was performed for this review. No queries were run, no API calls were made, no accounts were created and no money was spent. The reading was done by an AI-assisted research browser, and I did not personally test these products. Source links go to the providers' official pages.

Known gaps: no answer differences were measured; the exact consumer model routing is not established beyond what the support pages name; consumer subscription prices were not compared; a location setting for Gemini API grounding was not found in the guide checked, with no absence claim; controls specific to AI Overviews personalization were not established; the terms are observed public versions, not legal advice or negotiated agreements; cloud reseller terms are out of scope. Perplexity was not assessed in this research pass.

Changelog: 8 November 2026, first version.

Sources