How to Audit AI Search Visibility: A Repeatable Website Test

An AI search visibility audit is a repeatable test of whether ChatGPT, Perplexity, and other answer engines can find your pages, use them, cite them, and represent current facts accurately. It is not a ranking forecast. Record the exact prompt, engine, answer, cited URL, and date, then fix the evidence gap you can verify. Perplexity says its answers include links to original sources, while OpenAI warns that search results and citations can be incomplete, outdated, or incorrect. Your audit should test the evidence, not trust a single screenshot. Read how Perplexity describes its search process and how ChatGPT recommends checking search sources.

What is an AI search visibility audit?

It is a dated evidence log for AI answers and the pages behind them. A useful audit connects a customer-style question to the answer an engine returned, the source URL it cited, and the page signals that support or weaken that source. The output is a short list of verified fixes, not a vanity score or a promise that an engine will mention you.

1. Freeze the test before you run it

The first step is to make the audit repeatable. Choose the engines, prompts, location, device context, and date before opening a search box. If you change the question every time you test, you cannot tell whether a later result reflects a site change or a different test.

1. Choose three to five questions a real buyer might ask, not only your company name.
2. Run the same wording in each engine and save the full answer, not just the first sentence.
3. Record the date, signed-in state, location settings, and whether web search was enabled.
4. Copy every cited URL and label whether the answer mentioned your brand, cited your page, and stated a current fact.
5. Repeat important prompts before making a conclusion. One answer is a sample, not a stable ranking position.

2. Use prompts that resemble buying questions

A brand-name prompt tests recognition. It does not tell you whether a prospective customer can discover or use your site for a decision. Build a small prompt set around the jobs your website is supposed to support.

• Category: What should someone look for when choosing a [service or product] in [market]?
• Problem: How can someone solve [specific problem] without [common trade-off]?
• Comparison: Which options are suitable for [use case], and how do they differ?
• Proof: Which providers publish evidence for [claim, credential, result, or policy]?
• Brand: What does [business name] offer, who is it for, and where does it operate?

Keep the wording neutral. Do not pack the prompt with facts that only your own site uses, and do not treat a favorable answer as proof that your content caused the result. The goal is to see what a buyer encounters and which sources an engine chooses to show.

3. Score the answer and the source separately

A brand can be mentioned without being cited, cited without supporting the claim, or cited accurately but only for a narrow question. Keep those outcomes separate so the fix matches the failure.

• Not mentioned: inspect business clarity, entity wording, and whether the relevant page is discoverable.
• Mentioned but not cited: look for a page that answers the question directly and can be linked as evidence.
• Cited but inaccurate: correct the cited page and find conflicting or older pages that still state the old fact.
• Cited and accurate: keep the source in your baseline and test a different customer question before changing it.
• Cited from an old URL: consolidate the content, redirect a retired page when appropriate, and update internal links.

4. Check the cited page, not just the AI answer

Open every cited URL and compare the exact claim with the page as a visitor sees it. Check the title, H1, first paragraph, pricing or service details, location, availability, and any date that could change the interpretation. If the answer is wrong, identify whether the source itself is wrong, an older page is competing with it, or the answer went beyond what the source supports.

Perplexity's own support guidance says its citations link to original sources and recommends checking those sources. It also accepts reports for misinformation and outdated information through the flag control or support channels. Use the cited-source workflow described by Perplexity when the problem is an answer error, but fix your own page when the source is the problem.

5. Verify crawl access and rendered content

A source can be correct for a human and still be difficult for a crawler to retrieve. Check the HTTP response, robots.txt, firewall or bot rules, canonical URL, and the raw HTML returned without running a browser. Compare that response with the text a visitor needs to see.

Perplexity documents PerplexityBot as the crawler that surfaces and links websites in its search results, and recommends allowing it when you want your pages to appear there. OpenAI says sites should avoid blocking OAI-SearchBot if they want content included in ChatGPT search summaries and snippets. These are eligibility checks, not inclusion guarantees: OpenAI explicitly says placement is not guaranteed. See the Perplexity crawler documentation and OpenAI's publisher guidance for the current vendor rules.

6. Check structured data and real dates

Structured data can make facts easier for machines to interpret, but it cannot repair a page that says something else. Compare Organization, LocalBusiness, Product, Article, or FAQPage JSON-LD with the visible content and remove fields you cannot support. Google's guidelines recommend JSON-LD, require structured data to represent the page, and warn against marking up content that is hidden or misleading. Use the Google structured data guidelines as the validation boundary.

Dates need the same discipline. Show a real publication or last-updated date when it helps a reader, and change it only after a substantive edit. Google's guidance says visible and structured dates should be consistent, should not be future dates, and do not guarantee that a date will appear in Search. A sitemap's lastmod can help Google understand significant updates, but Google calls a sitemap a hint rather than a guarantee of immediate crawling. See Google's date guidance and sitemap documentation.

If several pages repeat the same business fact, choose one canonical source and link to it consistently. Retire or redirect obsolete duplicates when that is appropriate for users. Google's canonicalization documentation explains why duplicate pages can make it harder to understand which URL represents the content. That is useful hygiene for search systems generally, but it is not a direct Perplexity control.

7. Turn observations into a prioritized fix list

Prioritize the smallest change that removes a verified failure. Do not rewrite every page because one answer changed. A practical order is access first, source truth second, direct-answer copy third, and supporting proof fourth.

1. Fix a blocked, broken, redirected, or empty page before optimizing its wording.
2. Correct the canonical page and remove contradictions from old pages, profiles, and downloadable documents you control.
3. Add the direct answer and the boundaries a buyer needs, then link to the supporting evidence.
4. Validate visible facts and JSON-LD together; do not add schema-only claims.
5. Save the original prompt and rerun it after the change so the next result is comparable.

Answer Patch's Answer Readiness, Technical Access, and Structured Data pages describe the three common fix areas in this workflow. The links are useful even if you are doing the audit yourself: each one gives you a narrower implementation target instead of a generic request to optimize for AI.

How often should you repeat an AI search visibility audit?

Run a baseline now, repeat the same prompts after a material change, and keep a schedule your team can maintain. Re-test when you change an offering, price, location, policy, canonical URL, robots.txt rule, or key answer page. A dated log is more useful than a one-time screenshot because it shows which claims changed, which sources changed, and which fixes still need verification.

What can't an AI search visibility audit prove?

It cannot prove that ChatGPT, Perplexity, Google AI Overviews, or another system will rank, cite, recommend, or summarize your site in the future. It cannot turn one observed answer into a universal score, and it cannot establish why a system selected a competitor unless the available evidence supports that conclusion. Treat the log as a diagnostic instrument: it tells you what happened under defined conditions and which site signals you can improve.