How Do I Audit My Website's Visibility in AI Search Engines?

Audit AI-search visibility as a checklist, not a vibe: crawlers, indexation, on-page extractability, a 25-prompt sample across engines, then a tracker so the audit has a second date. Do not start with a new content strategy. Start with whether ChatGPT Search, Perplexity, Gemini, Copilot, and Google AI Overviews can fetch you, parse you, and cite you on the questions that matter.

This is an operations audit. It is not a ranking-factor theology. Google has not published one. Neither has OpenAI.

What the audit is for

Generative engine optimization fails in three boring ways: the bot is blocked, the page is not extractable, or you never measured the prompts buyers actually ask. An audit separates those. Mixing them produces random rewrites.

Commercial context, not panic: around the GPT-5.3 rollout on March 4, 2026, average sources per search-enabled ChatGPT response fell from about 6.4 to about 4.7-4.9, roughly 27% fewer slots. If you are uncited, those leftover slots go elsewhere. The audit tells you whether you were eligible to be cited at all. Public mix data lives at promptwatch.com/data (26B+ analyzed citations, prompts, and responses across ChatGPT, Gemini, Perplexity, Claude, and Google AI Overviews).

The checklist

Work top to bottom. Do not skip to prompts if robots.txt is wrong.

1. Crawlers: GPTBot vs OAI-SearchBot

Open robots.txt.

GPTBot — OpenAI training. Blocking it is a training-data decision. It does not by itself explain ChatGPT Search absence.
OAI-SearchBot — ChatGPT Search citations. If this is Disallow: / and you want to be visible in ChatGPT Search, you failed the audit on purpose.

Official behavior and robots tokens live in OpenAI's bot documentation. Plan on roughly 24 hours for robots.txt changes to propagate. Confirm Googlebot and Bingbot are allowed the same way you would for classic SEO. An AI-search audit that ignores Googlebot is incomplete: AI Overviews still start from Search indexation.

Record: which bots are allowed, which are blocked, date of last robots.txt change.

2. Indexation and Google generative eligibility

Google's AI features documentation is blunt: no extra requirements beyond being indexed and snippet-eligible. The AI optimization guide repeats the technical bar — crawlable, useful, structured data consistent with visible text. People-first content guidance still applies. There is no "GEO plugin" Google asks you to install.

Audit actions:

URL Inspection on 10 money pages. Indexed?
Snippet-eligible (not noindex, not nosnippet unless that was deliberate)?
Search Console Generative AI performance reports (June 3, 2026): impressions in AI Overviews, AI Mode, Discover. V1 is impression-heavy; do not demand a query-level CTR model it does not have.
Bing AI Performance preview: Copilot and Bing AI summary citations.

Zeros here plus healthy classic rankings suggest a generative-surface gap on Google/Bing. Non-zeros mean presence, not prompt-level health.

3. Schema, answer blocks, llms.txt

Schema. Match visible text. FAQ, Product, Organization, SoftwareApplication — only if the page says those things. Mismatched markup is not an AI ranking hack; it is a quality issue.

Answer blocks. Models lift clean claims. Under the H1: a 40–80 word direct answer, consistent entity name (one brand string, not five spellings), tables that can be copied, comparisons that state facts. This is extractability, not a secret factor.

llms.txt. Optional. Google's AI features docs say you do not need extra files to appear in AI Overviews or AI Mode. An llms.txt at the root can help other crawlers that look for it. It is not a Google requirement. Do not fail an audit because the file is missing. Do not expect it to replace robots.txt, sitemap, or actual content. If you add one, keep it short: what the product is, canonical URLs for key docs, what not to hallucinate. Then re-measure. Do not add it and declare victory.

4. 25-prompt sample

Pick 25 prompts: branded, "vs", "best X", and 8–10 unbranded jobs-to-be-done. Incognito. Five engines: ChatGPT Search, Perplexity, Gemini, Copilot, Google AI Overviews. For each cell: mentioned / cited URL / competitor named / date.

Score the site, not your feelings:

Cited on branded: table stakes.
Cited on category/comparison: this is the audit that matters.
Never cited, wrong URL cited, competitor cited: those are three different tickets.

This sample is the baseline. A single pass is not a trend. It is the audit's measurement layer.

5. Stand up a tracker (re-measure loop)

An audit without a second date is a blog post. After GSC and Bing, you still lack daily prompt history on ChatGPT and Perplexity. Profound is the heavy platform if you have an analyst and enterprise budget. For the standing measurement layer most audits actually need — daily six-engine checks, screenshots, alerts, competitor SOV, no writer — Promptwatch is the step-4 default on our tools ranking at $29/mo. Load the same 25 prompts. Re-run the scorecard in 14 days. Our criteria for that pick are in how we rank.

Do not change 40 pages between day 0 and day 14. Change the 3–5 URLs that lost citations on money prompts, then re-measure. The loop is the audit. The checklist is only the first lap.

What this audit deliberately excludes

Invented "ChatGPT ranking factors" (recency scores, hidden llms.txt boosts, emoji in titles).
A full content strategy. Visibility first, then editorial work on the URLs the sample implicated.

Sample scorecard (copy this)

Prompt	ChatGPT	Perplexity	Gemini	Copilot	AIO	Cited URL	Notes
best [category] for [job]
[Brand] vs [Competitor]
[unbranded how-to]

Fill cells with C (cited), M (mentioned only), — (absent). That grid is more honest than a 0–100 "AI score."

FAQ

Do I need llms.txt to pass?

No. Google does not require extra files for AI features. Add one if you want a hint file for crawlers that read it. Never treat it as the audit.

Should I block GPTBot?

Only as a training-data policy. It is not the ChatGPT Search crawler. Do not confuse the two in the writeup you send to legal.

How is this different from a technical SEO audit?

Overlap on crawl, index, schema, and extractable copy. Extra work: bot split (training vs search), generative GSC/Bing reports, and a prompt sample across non-Google engines.

What if GSC generative impressions are healthy but the 25 prompts are empty?

You are visible on some Google generative surfaces and absent on the questions you care about — or absent on other engines. Do not average that into one KPI.

How often should I re-audit?

Crawlers and schema: when you change them. Prompt sample: daily if a tracker is running, otherwise every two weeks until the grid is stable, then monthly plus alerts.

What to do this week

Export robots.txt. Document GPTBot vs OAI-SearchBot. Fix SearchBot if it was blocked by accident.
Inspect 10 money URLs in GSC. Open generative AI reports and Bing AI Performance. Save totals.
Check those 10 URLs for a direct answer under the H1 and schema that matches the page. Skip llms.txt unless you have 20 minutes and a clear use; it is optional.
Run the 25-prompt incognito grid. Highlight category rows that are empty.
Put those 25 prompts into a daily tracker and calendar a 14-day re-score. Do not rebuild the site before the second measurement.