Skip to content

ChatGPT search optimization: how to get your business cited in ChatGPT and Google AI answers

13 Nov 2026

·

8 minute read

ChatGPT search optimization, like AI search optimization in general, is mostly three unglamorous things: allowing the crawler that controls citation, keeping your pages eligible to show a snippet, and making sure your structured data describes what a reader can actually see. All three are documented by OpenAI and Google in plain terms. The reason so much advice on this topic is vague is that it skips the documentation and guesses at the ranking behaviour instead, and the documented parts are the parts you control.

Here is what the two sets of documentation actually say, what OpenAI's crawlers did on our own site in one week, and what to change on your own site this week.

Blocking GPTBot does not remove you from ChatGPT's answers

This is the mistake worth fixing first, because a lot of sites made it in good faith.

OpenAI documents four separate crawlers, and each one is controlled independently:

  • OAI-SearchBot surfaces sites in ChatGPT's search features. This is the one that governs whether you can be cited.
  • GPTBot crawls content that may be used to train OpenAI's generative AI foundation models.
  • ChatGPT-User covers user-initiated visits from ChatGPT, Custom GPTs and GPT Actions. OpenAI states it "does not crawl automatically", and that robots.txt rules may not apply because a person triggered the request.
  • OAI-AdsBot checks the safety of landing pages submitted as ads, and visits only those pages.

When the first wave of AI-crawler blocking happened, a great many sites added a GPTBot disallow rule and considered the matter closed. That rule is about training. It says nothing about whether ChatGPT can cite you in an answer, which is OAI-SearchBot. A company can therefore be absent from a training corpus and still cited daily, or blocked from citation while its content trains models, depending on which token it picked.

The settings being independent is the useful part, because it lets you take a position rather than an accident:

Two groups in robots.txt do it: User-agent: OAI-SearchBot followed by Allow: /, and User-agent: GPTBot followed by Disallow: /. That combination says: cite me, do not train on me. The reverse is equally expressible. OpenAI notes that when both are allowed it may use a single crawl for both purposes, that blocked sites can still appear as navigational links, and that robots.txt changes take around 24 hours to affect search results.

Go and read your own robots.txt before you read any further. Most teams do not know which of those four tokens they currently allow, and the answer is usually an accident of a template or a plugin default rather than a decision anyone made.

Google's bar is ordinary indexing, not a new discipline

The AI-SEO industry has produced a great deal of advice about how to earn a place in AI Overviews. Google's own documentation on AI features in Search is much duller than that advice, and duller is better here, because it is checkable.

To be a supporting link, a page must be "indexed and eligible to be shown in Google Search with a snippet". Google adds that there are "no additional technical requirements". It also says plainly that meeting the requirements guarantees nothing, since "indexing and serving isn't guaranteed".

Two practical consequences follow from that one sentence about snippets.

If you have set nosnippet, or a restrictive max-snippet, or wrapped sections in data-nosnippet, you have limited what can be shown from your pages, and that is the same eligibility AI features depend on. Some sites added those tags during the same defensive wave as the GPTBot rule, without connecting the two. Check them.

Separately, Google-Extended governs training and grounding in some of Google's other systems, and is not the control for Search. Treat it as its own decision, the way GPTBot is separate from OAI-SearchBot.

Beyond that, Google's recommendations are the ordinary ones: "creating helpful, reliable, people-first content", keeping important content available as text rather than locked inside images, and keeping business information current. Text matters more than it used to. A specification that exists only in a product photo, or a price that lives in a PDF, is not available to be quoted.

When you change any of these controls, expect latency. Google's own guidance on checking preview settings notes that recrawling can take anywhere from several days to several months.

Structured data has to match what the reader sees

Structured data helps a machine read a page it would otherwise have to infer. Google's introduction to structured data recommends JSON-LD, while noting that "all 3 formats are equally fine for Google, as long as the markup is valid and properly implemented".

The constraint that matters is honesty. Google states: "don't add structured data about information that is not visible to the user, even if the information is accurate." The accuracy clause is the interesting half. Marking up a rating, a price or an opening hour that is true but not shown on the page is still outside the guidance. It also warns against building pages that exist only to carry markup.

The consequence is softer than people assume. Google does not describe a penalty for this, it says the structured data "might be ineligible for rich result display in Google Search". The cost is the thing you were trying to gain.

For a small business the useful subset is small: who you are, what you sell, where you operate, and how someone contacts you, marked up and also written on the page in words.

What OpenAI's crawlers did on our own site in one week

We checked our own site in the hosting provider's AI crawler report for the seven days to 17 November 2026. Our robots.txt has one rule for every crawler, allowing everything except the admin area, and the provider's per-crawler block switches were all off.

OpenAI crawlerWhat it is forSuccessful requestsOther responses
GPTBotTraining40828
ChatGPT-UserVisits a person asked ChatGPT to make1511
OAI-SearchBotChatGPT search, and so citation37

The crawler that decides whether ChatGPT can cite a page made three successful requests in a week, against 408 from the training crawler. On a small site each search-crawler visit is scarce, so it should land on a page that answers at once with the content, without a redirect chain or a moved page that has no redirect. "Other responses" in the report means anything other than success, which includes redirects and not-found pages. When we looked at the not-found responses from all AI crawlers over one day, the most common were scanners asking for configuration files, which should fail.

Two lessons came out of it. A user-agent test proves less than it seems: a request from our own machine carrying OAI-SearchBot's user-agent string got a normal page, but hosts recognise the real crawler by the IP ranges OpenAI publishes, so the host's own report is the better evidence. And that report is the place to look after any site move or redirect change, because it shows whether the crawlers that matter are getting pages or errors.

What to change this week

  1. Open your robots.txt and write down which of OAI-SearchBot, GPTBot, ChatGPT-User and Google-Extended you currently allow. Treat whatever you find as an accident until someone confirms otherwise.
  2. Decide citation and training separately, because they are separate switches. Allowing citation while refusing training is a coherent position, and so is the opposite.
  3. Search your templates for nosnippet, max-snippet and data-nosnippet. Each one narrows the snippet eligibility that Google's AI features sit on top of.
  4. Move the facts you want quoted into text. Prices, specifications, hours and contact details should appear in the HTML itself, where a crawler can read them. An image or a PDF attachment keeps them hidden.
  5. Check that your structured data describes what is visibly on the page, and remove anything that is true but not shown.
  6. Allow for the delay. OpenAI says about 24 hours for robots.txt changes to reach search results. Google says a recrawl can take days to months.
  7. If your host or CDN reports crawler traffic, check that OAI-SearchBot and Googlebot get successful responses, especially after a site move or a change to redirects.

None of this makes a model mention you. It decides whether it is allowed to, and whether it can read you accurately when it does. That is the part under your control, and it is worth getting right before spending anything on the part that is not.

The same documentation decides what an AI agent sees when it fetches one of your pages on a user's behalf, which is a related problem we have written about in what safeguards should control an AI agent in production.

If you want someone to go through your crawler rules and structured data properly, tell us about the project.

AI tools

Ask ChatGPTAsk Claude

Start with the problem that matters