Skip to content

Guides

Is scraping ChatGPT legal? What the terms of service say

OpenAI, Microsoft and Anthropic bar automated extraction from their chat interfaces, and the official routes add limits: Gemini grounding caps storage at two years.

A blue card slides out from a stack of white layered blocks on a light background

SEO consultant, and more recently GEO, at MentionBox

Collecting ChatGPT answers with a script breaks OpenAI's Terms of Use, which list "Automatically or programmatically extract data or Output" among the things you cannot do. Microsoft writes the same rule for its AI services, Anthropic bars crawling and scraping in its consumer terms, and Google states it narrowly, forbidding "using automated means to access content from any of our services in violation of the machine-readable instructions on our web pages" in Terms of Service effective July 30, 2026. None of that makes the act a crime in the United States: the Ninth Circuit held on April 18, 2022 that the Computer Fraud and Abuse Act's "without authorization" concept "does not apply to public websites". The exposure is contractual, and it lands on the account and the key you collect with.

This page is for whoever is about to buy or build an AI visibility feed and wants to read the governing documents first. We are not lawyers and none of this is legal advice: the clauses are quoted so you can take them to counsel with the source next to them.

What each engine's own terms say

EngineDocument governing the interfaceWhat it says about automated collectionOfficial route and its price
ChatGPTOpenAI Terms of Usebars automated or programmatic extraction of data or Outputweb search tool on the Responses API, $10 per 1,000 calls plus content tokens
Google AI Mode and AI OverviewsGoogle Terms of Service, Search spam policiesautomated access in breach of machine-readable instructions; rank-checking named as machine-generated trafficnone for the AI answers; Gemini grounding is a separate surface, $14 per 1,000 searches
Microsoft CopilotCopilot Supplemental Terms"Don't use tools or computer programs (like bots or scrapers) to access Copilot."Grounding with Bing Search, whose terms forbid storing Output
ClaudeAnthropic Consumer Termsbars crawling and scraping, and automated access without an API keyClaude API with the web search tool
Perplexitylegal pages unreachable from our networknothing quoted hereSearch API, $5.00 per 1,000 successful requests

Clauses and prices quoted from each company's own terms, policies and developer documentation, read on November 8, 2026. Perplexity's legal pages returned an access error, so this page makes no claim about their contents.

Three of the five write a flat prohibition on automated extraction from the consumer product and point developers at a paid API. Google writes a conditional one, tied to its robots directives, and puts its sharpest wording in a different document.

OpenAI: the Terms of Use and the robots file

OpenAI's prohibition sits in the list of things you cannot do, next to reverse engineering the Services, and it covers data and Output, which is the whole payload an answer-tracking pipeline wants.

The robots file is the more concrete document. chatgpt.com/robots.txt opens with an allowlist for User-agent: *, naming the exact homepage with Allow: /$, the public share paths, the marketing pages and about sixty locale homepages, then closes with Disallow: / under the comment "Now block everything else". Two redundant lines underneath decide the question for an interface collector:

# after the allowlist and the blanket Disallow: /
Disallow: /backend-anon/sentinel/*
Disallow: /backend-anon/conversation$

Those are the paths an unofficial collector hits, and the same file blocks named AI crawlers outright, including CCBot, Google-Extended, anthropic-ai and PerplexityBot. By comparison openai.com/robots.txt is 98 characters of Allow: /, so the restriction targets the product.

The sanctioned route carries a condition rather than a storage ban. OpenAI's web search tool guide states that when your interface displays web results or information from them, "you must show inline citations as prominent, clickable links", at $10 per 1,000 calls with search content tokens billed at model rates. Our comparison of the OpenAI web search tool and ChatGPT answers sets out where the two surfaces diverge.

Google: automated queries to Search and to AI Mode

The word "scraping" appears once in Google's Terms of Service, in the suspension clause, which covers conduct causing harm or liability "by hacking, phishing, harassing, spamming, misleading others, or scraping content that doesn't belong to you". One bullet under "Don't abuse our services" is unconditional: using "AI-generated content from our services to develop machine learning models or related AI technology" is prohibited outright.

The flattest statement lives in the Search documentation. Under the heading "Machine-generated traffic", Google's spam policies say the practice "refers to the practice of sending automated queries to Google", that "this includes scraping results for rank-checking purposes or other types of automated access to Google Search conducted without express permission", and that "such activities violate our spam policies and the Google Terms of Service". Sites that breach the policies "may rank lower in results or not appear in results at all". The help page on automated traffic lists the sources it means, from search scrapers to "use of software that sends searches to Google to find how a website or web page ranks on Google", with a reCAPTCHA interstitial and network-level blocking as the remedies.

There is no official API for the AI answers. Google's guidance on AI features and your website points site owners to the Search Console Performance report, where AI Overviews and AI Mode traffic is "included in the overall search traffic" inside the "Web" search type, with no separate dimension. The Custom Search JSON API is not a substitute: it queries a Programmable Search Engine you configure, and it is "closed to new customers", with existing customers given "until January 1, 2027 to transition". Our guide on getting Google AI Mode answers through an API prices the vendors who fill that gap.

The Gemini API does not lift those limits

Teams often assume the paid Google route settles the compliance question. Its additional terms, effective March 23, 2026, read the other way for a measurement use case. The grounding section states that you "will not, and will not allow your end user or any third party to, cache, frame, syndicate, resell, analyze, train on, or otherwise learn from Grounded Results or Search Suggestions", and adds that "it is a violation of these terms to use Grounding with Google Search to extract or collect one or more of these components for another purpose", the examples given being programmatic collection of links, index building and crawling.

Three further clauses shape what a pipeline may do:

  • A narrow storage exception. You "may copy and store, for up to two (2) years, the text of the Grounded Result(s)", limited to evaluating display, end-user chat history and temporary resubmission through a function call.
  • A display obligation. Grounded results go only to the end user who sent the prompt, with their Search Suggestions, five at most, unmodified.
  • Survival. The subsection survives termination, so closing the project does not release results already stored.

Billing is per search the model decides to run: $14 per 1,000 search requests after 5,000 free a month shared across Gemini 3 and newer models, with a default limit of 1,500 requests a day per project, on the grounding page. The general Google APIs Terms of Service, last modified November 9, 2021, say it more widely at section 5(e)(1): you will not "scrape, build databases, or otherwise create permanent copies of such content, or keep cached copies longer than permitted by the cache header". Our explainer on grounding with Google Search in the Gemini API walks through a request.

Microsoft and Anthropic: the plainest wording of the five

Microsoft states the consumer rule in one line. Its Copilot Supplemental Terms of Use, effective August 18, 2026: "Don't use tools or computer programs (like bots or scrapers) to access Copilot." The same document adds "You can only use Copilot for your own personal use." The Microsoft Terms of Use carry a clause headed Extracting Data for AI services: "You may not use web scraping, web harvesting, or web data extraction methods to extract data from the AI services."

The official Bing route is the strictest document here for a visibility product. The Grounding with Bing Search use and display requirements, last updated November 2025, forbid you to "Copy, store, cache, archive, or create a database of Output" at section 4(b)(o), require references to be displayed "in the exact form provided by Microsoft", and state at section 2(e) that "you may not use any Output to train any AI Models". A pipeline whose purpose is a time series of answers cannot run on that route.

Anthropic's consumer terms, in the European edition effective October 8, 2025, prohibit at section 3 "to crawl, scrape, or otherwise harvest data or information from our Services other than as permitted under these Terms", and bar access "through automated or non-human means, whether through a bot, script, or otherwise" without an API key. Its Usage Policy, in the version marked effective November 12, 2026, closes the workaround: do not "resell, proxy, or otherwise provide access to Claude through unauthorized means", with a qualifier naming "services that route requests through consumer subscriptions". That explains a visible gap in the market, since Claude is one of the engines fewest collection APIs reach. Our ranking of AI answer scraping APIs shows which engines each key reaches today.

What courts have decided

DecisionCourt and dateWhat it settled
Van Buren v. United StatesSupreme Court, June 3, 2021, 6 to 3the CFAA "does not cover those who ... have improper motives for obtaining information that is otherwise available to them"
hiQ Labs v. LinkedInNinth Circuit, April 18, 2022"the concept of 'without authorization' does not apply to public websites", and letting a company decide alone who may collect data it publishes "risks the possible creation of information monopolies"
Meta Platforms v. Bright DataN.D. Cal., Judge Edward M. Chen, January 23, 2024granted Bright Data summary judgment on breach of contract, after analysing "whether the Terms, properly construed, bar logged-off automated scraping of publicly available data"

Opinions read on the courts' own publication sites and on Cornell's Legal Information Institute, November 8, 2026.

Together they say that reaching a page open to anyone with a browser is not the federal access crime, and that whether a contract binds a logged-out collector depends on how that contract was drafted. The Meta ruling turned on the wording of one set of terms, so it is no licence to ignore another company's.

A fourth matter shows where the pressure moved. Google announced a suit against the SERP vendor SerpApi on December 19, 2025, built on section 1201 of the Digital Millennium Copyright Act rather than on access law, arguing that its anti-scraping system, named SearchGuard in the complaint, is a technological measure the vendor circumvented. The complaint estimates the vendor "sends hundreds of millions of artificial search requests each day to Google", and the court granted the vendor's motion to dismiss on July 21, 2026, reported by Search Engine Roundtable. Copyright rather than access law is now the live theory, and a complaint dismissed at that stage settles nothing about the practice.

What European rules add

The Court of Justice settled the contract point in Ryanair v PR Aviation, C-30/14, on 15 January 2015. Where a database is protected neither by copyright nor by the sui generis right, the Database Directive "is not applicable", so its mandatory user rights "do not preclude the author of such a database from laying down contractual limitations on its use by third parties". The terms at issue banned exactly this article's subject, screen scraping for commercial purposes without a written licence. Weaker intellectual property in the data leaves the terms of use standing.

The text and data mining exception in Directive (EU) 2019/790 of 17 April 2019 cuts the other way, within limits. Article 4 requires Member States to allow reproductions and extractions "of lawfully accessible works and other subject matter" for mining, and Article 4(2) lets you keep those copies "for as long as is necessary". Article 4(3) sets the condition: the exception applies only where the use "has not been expressly reserved by their rightholders in an appropriate manner", and for content published online the recitals point to "machine-readable means" as the appropriate form. A robots.txt that disallows your collector is that reservation.

The three routes and their limits

RoutePublished price per 1,000 AI answersWhat the governing document allowsWhat it costs in practice
GEO metrics API, for example API for GEO$13, down to $7 at its largest tierits own terms name resale and white label as an intended usethe highest price per answer of the three
AI answer collection API, for example DataForSEO or SerpApi$1.20 and $25the vendor's contract with you, which is not the engine's contract with the vendorengine coverage that moves, resale rights that differ by vendor
Official engine API$10 per 1,000 OpenAI calls, $14 per 1,000 Gemini searches, $5.00 per 1,000 Perplexity Search requestsdisplay and storage conditions, with caching and database building prohibited on both grounding routesthe answer a developer gets, not the answer a buyer sees

Vendor prices from our October 23, 2026 price check; engine prices from each company's own pricing pages, read on November 8, 2026.

The third row is the one that surprises people. The official route is the compliant way to run a model and the least usable way to measure a brand over time, since storing the result is what the grounding terms forbid. API for GEO is the example in the first row because its terms treat resale of the numbers as intended use. MentionBox, which publishes this site, has a commercial link with API for GEO.

Six mistakes teams make

  1. Reading one document. The governing text is usually spread over three: the terms of use, an acceptable use or spam policy, and the robots file. Google's hardest sentence on rank-checking is in the spam policies.
  2. Assuming the paid API is the permissive route. Bing grounding forbids copying, storing, caching or archiving Output, and the Gemini terms forbid caching or analysing grounded results outside the two-year exception.
  3. Treating a vendor's terms as the engine's terms. An engine tightening its defences reaches you as an outage or a new rate card, whatever your vendor contract says.
  4. Reading a logged-out win as a general rule. The Meta ruling construed one set of terms, and Microsoft's Copilot wording bars bots outright.
  5. Ignoring the output-side clauses. Google prohibits using AI-generated content from its services to develop machine learning models, and Anthropic requires prior authorisation before training on outputs.
  6. Promising a Google AI Mode history without checking the feed. There is no official API for those answers, and Search Console reports them inside the aggregate "Web" search type.

Our view

In our view the interfaces are closed by contract and the official APIs are closed by use restriction, which leaves most of this market running on the vendor's risk rather than yours. Ask a vendor which engines it collects from the interface, which from an official API, and what its contract says about resale. Our ranking of AI visibility APIs scores 55 keys on price, engines and data returned, and our explainer on query fan-out covers the field that shows what an engine searched before answering.

Frequently asked questions

Is it legal to scrape ChatGPT answers?

In the United States it is a contract question rather than a criminal one. OpenAI's Terms of Use list automated or programmatic extraction of data or Output among the things you cannot do, so collecting answers from the interface breaks the agreement you accepted. The Ninth Circuit held on April 18, 2022 that the Computer Fraud and Abuse Act's "without authorization" concept "does not apply to public websites", and chatgpt.com's robots.txt closes the conversation paths to compliant crawlers while leaving the homepage and public share links open.

Does Google allow automated queries to AI Mode or AI Overviews?

No. Google's spam policies for web search state that machine-generated traffic "includes scraping results for rank-checking purposes or other types of automated access to Google Search conducted without express permission", and that "such activities violate our spam policies and the Google Terms of Service". The Terms of Service, effective July 30, 2026, list automated access that violates the machine-readable instructions on Google's pages among the ways you must not abuse its services, and name scraping content that does not belong to you as a ground for suspension.

Is there an official API for Google AI Mode or AI Overviews?

Not for the answers themselves. Google's guidance for site owners points only to the Search Console Performance report, where AI Overviews and AI Mode are counted inside the overall "Web" search type with no separate dimension. The Custom Search JSON API searches a Programmable Search Engine you configure rather than general Google Search, is "closed to new customers", and gives existing customers until January 1, 2027 to move, at 100 free queries a day and $5 per 1,000 beyond that.

Can I store the answers an official API returns?

It depends on the route, and two of them say no. The Gemini API's additional terms, effective March 23, 2026, forbid you to "cache, frame, syndicate, resell, analyze, train on, or otherwise learn from Grounded Results or Search Suggestions", with a narrow exception allowing you to copy and store the text of a grounded result for up to two years. Microsoft's Grounding with Bing Search terms, last updated November 2025, prohibit you from "Copy, store, cache, archive, or create a database of Output". OpenAI sets a display obligation instead: "you must show inline citations as prominent, clickable links".

Do the terms still apply if I never log in?

That is the question a US court answered in a related case. On January 23, 2024 Judge Edward M. Chen granted Bright Data summary judgment on Meta's breach of contract claims, holding that Meta's terms, properly construed, do not bar logged-off automated scraping of publicly available data. The ruling turns on how those particular terms were drafted, not on a general rule, and the Microsoft Copilot supplemental terms show the other drafting style: "Don't use tools or computer programs (like bots or scrapers) to access Copilot."

Why do most AI scraping APIs not cover Claude?

Because Anthropic's documents close both doors. Its consumer terms of service bar you "To crawl, scrape, or otherwise harvest data or information from our Services other than as permitted under these Terms", and bar automated or non-human access by bot or script unless you hold an API key. Its Usage Policy, in the version marked effective November 12, 2026, adds a prohibition on reselling, proxying "or otherwise provide access to Claude through unauthorized means", naming services that route requests through consumer subscriptions.