How to check if my site appears in AI search

What counts as your site appearing in AI search?

Your site appears as a source when an observed answer links to a page on your domain. A brand mention without that link is a different visibility outcome.

Begin with three separate checks: whether the brand is named, whether your domain is cited, and whether the cited page supports the relevant statement. These outcomes often overlap, but none implies all the others.

A software brand may be recommended using third-party documentation, while an educational page may be cited without any recommendation of its publisher. Your checking sheet should preserve that distinction from the first observation.

Define the domain boundary before testing. Decide whether documentation subdomains, regional sites, an acquired product domain, and a public knowledge base count as owned sources. Record external company profiles separately.

A LinkedIn article written by your founder may be useful evidence, but it is not a citation to your website. Consistent ownership rules prevent different reviewers from reaching different conclusions about the same answer.

For a hypothetical workflow automation vendor, an answer that names the product but links only to an independent comparison earns a brand mention, not an owned-site citation.

An answer linking to the vendor’s connector documentation earns an owned-site citation even if it never recommends buying the product. Neither observation alone establishes qualified demand. This guide is a practical appearance audit, not a traffic attribution model.

Create an explicit result vocabulary before opening the engines. Useful labels include observed owned citation, observed brand mention only, no observed inclusion, and inconclusive test.

The last label matters when search was unavailable or the interface did not expose enough evidence. Avoid converting an incomplete test into a definitive statement that your site is missing from every AI search experience.

Which search surfaces should you check separately?

Test the actual interfaces your buyers use and record each one separately. Do not assume that ChatGPT, Google AI Mode, Google AI Overviews, or another assistant shares a single searchable index or reporting system.

Choose a manageable initial set of surfaces based on your audience and available access.

Record the product name, interface, displayed model information where available, account state, date, language, and relevant location settings. Do not invent unavailable model identifiers.

If the interface changes during your audit, annotate the change so that later reviewers understand which experience produced each answer.

OpenAI explains that ChatGPT can search the web automatically and that users can also initiate search. For a search appearance check, record whether you actually observed search and preserve the response’s source links.

A conversational answer without visible retrieval evidence should not silently enter the same bucket as a searched answer.

Its wording may be informative, but the test answers a different question. [1](https://help.openai.com/en/articles/9237897-searching-the-web-with-chatgpt)

Google documents AI Overviews and AI Mode as Search experiences that can surface relevant supporting links. It also explains that AI features may use related searches to develop a response.

Check the specific experience you can observe rather than assuming that the same wording must produce identical source selections in both.

If a query does not trigger an AI Overview, record the absence of that feature, not a failed citation opportunity. [2](https://developers.google.com/search/docs/appearance/ai-features)

Keep API experiments separate from consumer interface observations. An application configured with domain restrictions or a custom retrieval source does not demonstrate unassisted discovery in a public assistant.

Similarly, a direct instruction to read your website is a retrieval diagnostic rather than a neutral discovery test.

The distinction makes the final result more useful because it identifies which part of the appearance problem you have actually examined.

Sources: OpenAI explains ChatGPT web search. · Google documents AI features and website eligibility.

How should you build a small discovery test?

Use realistic unbranded buying questions first, then add branded and direct-page checks as separate diagnostic cohorts. Freeze the wording before comparing results.

Start with customer evidence such as recurring sales questions, implementation concerns, and public support issues that match your product category. Remove private customer information before using it in an external tool.

Your initial cohort should cover a few important needs in depth rather than every imaginable keyword. Label each prompt with the buyer problem, relevant capability, and reason the company might reasonably appear.

A hypothetical category prompt could state that a buyer needs subscription analytics software that works with a specified billing workflow and supports a documented reporting requirement.

Include only requirements that genuinely matter to that buyer. Do not insert your brand, an unusual slogan, or a uniquely identifying feature combination solely to induce inclusion.

That would measure a guided identification exercise rather than ordinary category discovery.

Use branded questions to test descriptive accuracy separately. A prompt might ask what a named product does, which integrations its documentation supports, and which limitations a buyer should verify.

A direct-page diagnostic can ask the assistant to inspect a particular public URL and summarise its stated prerequisites.

Those checks can uncover interpretation or retrieval issues, but they must not inflate the unbranded discovery result.

Save a versioned prompt document before running the audit. Include exact wording, cohort labels, inclusion criteria, and any exclusions.

If you later discover an important missing use case, add it to an exploratory cohort and retain the original baseline. Otherwise, improvements may reflect easier questions rather than better discovery.

The goal of this initial check is a trustworthy picture of selected buying situations, not a universal market share estimate.

How do you run each check without contaminating it?

Use a fresh conversation, preserve the exact prompt, and capture the full response with source links. Keep prompting conditions consistent enough that the records can be reviewed later.

Avoid continuing a conversation in which you have already introduced your company. Prior discussion can make a brand appear for reasons unrelated to discovery.

Use a fresh conversation for each independent run and record relevant personalisation settings where the product exposes them.

A fresh conversation does not prove that every influence has disappeared, so describe the conditions you controlled rather than claiming a perfectly neutral environment.

Run the agreed prompt without repeatedly nudging the answer toward your website. If the first answer omits your brand, save it before doing any diagnostic follow-up.

Asking why the company was omitted can be useful exploration, but the resulting explanation is not a reliable disclosure of the engine’s internal ranking process.

Treat the explanation as generated text that may suggest hypotheses, not as authoritative evidence of causation.

Record the run status, complete answer, visible citations, timestamp, and any observable search activity. Save enough context to distinguish a product mention in a comparison from a source link in a supporting panel.

Where export tools are unavailable, retain screenshots and a structured transcript together. Links can change, so copying only the final URL without the supported statement may make later review unnecessarily difficult.

Repeat a subset of the tests across separate runs or observation windows when practical. OpenAI’s evaluation guidance notes that generative outputs can vary for the same input.

Repetition helps reveal whether an appearance is persistent in your sample or occurred only once.

It does not establish that your small sample represents all users, locations, or conversations, and it should not be described that way. [3](https://developers.openai.com/api/docs/guides/evaluation-best-practices)

Sources: OpenAI explains evaluation design and output variability.

How do you verify a citation instead of trusting its label?

Open the cited destination and check its domain, content, and relationship to the generated claim. A plausible citation label is not enough to establish support.

Expand the available source controls and capture the destination URL. Note whether the link goes to your primary website, a documentation subdomain, a syndicated copy, or a third-party page.

If a redirect occurs, record the final destination and the original link where feasible. Apply the ownership rules you established earlier rather than deciding ownership ad hoc based on whether the answer appears favourable.

Read the passage that appears to support the generated statement. Confirm whether the page actually documents the capability, limitation, or comparison attributed to it.

A pricing page might be relevant to a commercial question but fail to support a specific integration claim. Record that as a support problem even when the site successfully earned a citation.

Visibility without factual support can create a worse buying experience than absence.

Distinguish a navigational link from substantive sourcing. A homepage link may help someone reach the company without supporting any detailed claim in the answer.

A source in a broader list may provide background rather than evidence for the sentence you are reviewing.

Where the interface does not make the relationship clear, mark support as uncertain and preserve the context instead of assigning a confident positive score.

Use a simple issue log with the generated statement, cited URL, relevant source passage, support judgment, and reviewer note. Escalate material inaccuracies to the product owner before changing public content.

The answer may be wrong because it combines sources incorrectly, because your source is outdated, or because an external page misstates the product.

Each explanation leads to a different action, so do not assume the website is always at fault.

What should you inspect when your site does not appear?

Review eligibility and access before assuming that more content is the answer. A missing citation in a small prompt sample does not establish that the site is blocked or excluded everywhere.

First confirm that the tested page is publicly reachable in an ordinary browser and that the information is not confined to a login wall.

Inspect status codes, redirects, indexability controls, canonical choices, and important text availability with an appropriate technical reviewer.

A successful browser visit is a useful check, but it does not prove that every crawler can retrieve the same content through your security layer.

For ChatGPT search, review OAI-SearchBot policy separately from GPTBot policy. OpenAI identifies these agents with different purposes and explains that search-related access involves both robots.txt and relevant network access.

Compare the documented user agent and published IP information with your delivery controls rather than trusting a user-agent string alone.

A spoofed crawler name is not sufficient evidence that a request came from OpenAI. [4](https://developers.openai.com/api/docs/bots)

For Google AI features, follow Google’s own eligibility and Search guidance.

Google states that supporting pages need to be indexed and eligible to appear in Search with a snippet, and that eligibility does not guarantee inclusion. Use Search Console and technical inspection to investigate specific pages.

Do not interpret a successful eligibility check as a promise that the next AI answer will select your page. [2](https://developers.google.com/search/docs/appearance/ai-features)

After access checks, review relevance and clarity. Your page may be available but not answer the tested buyer’s question, or it may bury the decisive requirement beneath broad positioning.

Improve the source where there is a genuine evidence gap, such as missing prerequisites or an outdated feature explanation.

Avoid creating near-duplicate pages for every prompt variation simply because a limited audit did not produce the desired citation.

Sources: OpenAI documents its search and training crawlers. · Google documents AI features and website eligibility.

How should you summarise the result of a first audit?

Report observed appearances, absences, and unresolved conditions separately. Include the sample definition so that readers know what the audit can and cannot establish.

Create an evidence table with a row for every completed test. Include the platform, prompt identifier, cohort, observation date, search status, mention status, owned citation status, cited URL, and accuracy note.

Keep technical failures visible in a separate field. A failed run is not an answer that omitted your site, and mixing the two can artificially depress your observed appearance rate.

At summary level, show counts alongside percentages. If your chosen audit contains only a handful of buying questions, readers should see that small base immediately.

Report the number of unique prompts as well as the number of completed answers when you repeat questions.

More repetitions of the same prompt increase observations of that scenario without automatically expanding the range of buyer needs covered.

Use restrained conclusion language. A defensible finding states that your domain appeared in selected searched answers under recorded conditions.

An equally useful negative finding states that no owned citation was observed in the completed unbranded tests. Neither statement should become a claim that the company is universally visible or entirely absent across AI search.

The limits are part of the result, not an optional footnote.

Attach a short action list with the evidence behind each recommendation.

Access failures belong with the technical owner, unsupported capability claims belong with product and content reviewers, and missing use-case detail belongs with the relevant page owner.

Prioritise issues by buyer impact and confidence in the diagnosis. A long list of speculative fixes can distract from a small number of verified problems that deserve immediate attention.

What should happen after the initial check?

Turn the audit into a limited correction and retest cycle. Preserve the baseline so that subsequent checks answer whether a defined problem changed rather than merely producing another screenshot.

Choose one or two problems with clear owners and factual remedies. If a page is inaccessible, resolve the access issue through the normal engineering process.

If a product statement is wrong, correct the approved source and note the publication date.

If the evidence is already accurate but the generated answer remains mistaken, record that fact rather than rewriting accurate documentation merely to resemble the model’s wording.

Retest the original questions under comparable conditions and retain all outcomes, including unfavourable answers. Record changes in the interface, prompt cohort, website, and competing sources that could affect interpretation.

A later citation is evidence of a later citation; it is not by itself proof that one particular edit caused it. Repeated observations can strengthen an operational judgment without eliminating every alternative explanation.

AEO Operators can help interpret an appearance audit and prioritise the evidence and access work behind it. Our free AI Search Grader is a one-time report on how ChatGPT, Perplexity, and Gemini describe a brand.

It can complement a manual site check, but it should not be described as continuous monitoring or a comprehensive census of every relevant AI answer. Match any service to the question you need answered.

Keep the first audit small enough to repeat. An elaborate baseline that nobody can maintain becomes an isolated document rather than an operating practice.

Once the checking method is reliable, you can expand the question coverage, add reviewers, and connect the findings to broader measurement.

The immediate objective remains straightforward: know where your site was observed, verify what the answer says, and make the next action defensible.

Frequently asked questions