GEO checklist: A practical AI search audit

A generative engine optimization (GEO) checklist audits whether your website can be found, read correctly and cited as a source in AI search. That covers ChatGPT search, Claude, Perplexity, Microsoft Copilot and Google's AI Overviews and AI Mode.

It covers five areas: scope, access, content, entity and brand facts, and measurement. Each check needs a piece of evidence and an owner. A list of ticked boxes does not show that your visibility improved.

Bot, report and setting details were checked against the vendors' own documentation on November 1, 2026. Where popular GEO advice contradicts it, the relevant section says so.

GEO (generative engine optimization) is the work of getting your content and brand into AI-generated answers, in the right context and with accurate facts. It differs from SEO in what you track. SEO looks at rankings and clicks. GEO looks inside the answer: is your brand mentioned, is your page linked as a source, and is what the answer says about you true?

Keep those outcomes apart. A brand can be named with no link, and a product recommended at the wrong price is visible in a way that hurts. That is why a GEO audit records the answer text itself.

The other difference is the number of platforms. ChatGPT, Claude and Perplexity run their own crawlers, with separate bots for search, training and user requests. That is where most new checklist items come from.

Are GEO, AEO and AI SEO different things?

Mostly they are different labels for the same job. AEO (answer engine optimization) stresses answer engines, AI SEO stresses the search channel, and GEO stresses generative engines. Google's AI optimization guide defines AEO and GEO and states that, from Google Search's perspective, optimizing for generative AI search is still SEO.

Is GEO replacing SEO?

Not for Google. Google says there are no additional requirements to appear in AI Overviews or AI Mode. A supporting link must be indexed and eligible to show with a snippet, with no other technical requirement and no guarantee of inclusion. Other platforms also have to crawl a page before citing it.

Canonical tags, sitemaps, redirects, JavaScript rendering and hreflang therefore stay prerequisites. Run the SEO checklist for them, and use the items below for what is specific to AI search.

A GEO checklist needs items from five areas: scope, access, content, entity and brand facts, and measurement. Use it on one priority page group at a time and record status, evidence, owner and next review date per item. If an item does not apply, write down why, so "not applicable" never looks like "not done".

1. Scope: List the target AI platforms, the priority pages and the questions customers are trying to answer. Keep branded and unbranded questions, and each market, separate. Evidence: a dated prompt set.
2. Scope: Save baseline answers with date, language, country and platform. Store the answer text, the cited URLs and any wrong facts in separate fields. Evidence: the baseline record.
3. Access: Confirm the classic crawl and index audit passes. Evidence: SEO audit output.
4. Access: Set robots.txt rules separately for search, training and user-initiated bots. Evidence: robots.txt plus a written decision.
5. Access: Record the Search generative AI control value in Search Console and whether it is inherited from a parent property. Evidence: a dated note of the setting.
6. Access: Check that the CDN and firewall let verified bots through, using the published IP lists. Evidence: server logs.
7. Access: Confirm every nosnippet, data-nosnippet and max-snippet rule is intentional. Evidence: page HTML and HTTP headers.
8. Content: Check that each priority page answers its main question early and states price, scope and exclusions as text. Evidence: page review.
9. Content: Give numerical claims a source, period and scope, and label illustrative calculations as examples. Evidence: source list.
10. Entity and brand: Make brand and product names, prices, service areas and contact details match across the site, Merchant Center and Business Profile. Structured data must match visible text. Evidence: comparison sheet.
11. Measurement: Record Google's Generative AI performance report and Bing AI Performance separately, with their definitions. Evidence: period report.
12. Measurement: Track AI referral visits and conversions, and log content changes by date and URL. Evidence: change log.

Do the scope items first, because they decide what the rest of the audit looks at. Start with pages that carry revenue or a customer decision. Branded prompts show how a known company is described but say little about discovery. So add category and use-case prompts, and keep each market's conditions in its own prompts.

How many prompts do I need, and how many runs per prompt?

None of the platform documentation reviewed for this article gives a number for either, so the answer comes from your decisions. Build one fixed set per priority page group and market and keep it unchanged between rounds.

Google notes that AI Mode and AI Overviews may use different models and techniques, so the responses and links they show will vary. Run each prompt more than once under the same conditions and report a rate.

As a hypothetical example, source links in six of twenty recorded answers mean a 30% linked-source rate for that observation group. They do not describe market share or what every user saw.

AI companies use separate bots for search, model training and user-initiated fetches, so decide access bot by bot. A single "block AI bots" switch ties search visibility to training consent.

Which AI bots should I allow?

If you want to be cited in AI search, allow the search bots: OAI-SearchBot, Claude-SearchBot and PerplexityBot. Decide on GPTBot, ClaudeBot and Google-Extended as a content-licensing question, because their vendors describe them in terms of model training. The table summarizes each bot's purpose and the effect of blocking it, based on the documentation from OpenAI, Anthropic, Perplexity and Google.

The pattern is consistent across vendors: appearing in AI search and allowing training are separate decisions. User-initiated agents are the exception to robots.txt control. OpenAI says robots.txt rules may not apply to ChatGPT-User, and Perplexity says Perplexity-User generally ignores them, while Anthropic states that its bots honor robots.txt.

Does blocking GPTBot hide me from ChatGPT search?

No. OpenAI says its settings are independent. It gives the example of allowing OAI-SearchBot to appear in search results while disallowing GPTBot to opt out of training. Blocking OAI-SearchBot is what removes a site from ChatGPT search answers.

OpenAI adds that a robots.txt change can take about 24 hours to reach its search systems. Anthropic splits ClaudeBot (training) and Claude-SearchBot (search) the same way. So guides claiming that blocking GPTBot or ClaudeBot removes all AI search visibility are not supported by the vendors' pages.

Google's three separate controls

On Google, three separate controls affect AI features. The first is the Search generative AI control in Search Console, rolled out to all websites worldwide as of August 31, 2026.

It decides whether your site can appear in AI Overviews, AI Mode and generative AI features in Discover. Include is the default, and child properties inherit from their parent unless an owner overrides it.

Google says exclusion generally takes a few days, with some content taking longer because of caching. The control does not affect AI training.

The second is Google-Extended, the robots.txt token for Gemini training and grounding covered in the table, which has no effect on Search.

The third is page-level snippet controls. Google's AI features documentation points to nosnippet, data-nosnippet, max-snippet and noindex for limiting what Search shows. A nosnippet rule also stops the content from being used as a direct input for AI Overviews and AI Mode. A forgotten nosnippet in a template can quietly remove pages from those features, so check the robots meta tag and the X-Robots-Tag header.

Verify CDN and bot protection against published IP ranges

A robots.txt allow rule means little if the CDN or firewall serves the bot a challenge page. Google lists crawling being allowed "by any CDN or hosting infrastructure" among the fundamentals for its AI features.

OpenAI and Perplexity publish a separate IP list for each bot. Perplexity advises matching each user agent against its own IP ranges, with an example Cloudflare rule. User-agent strings can be spoofed, so a name in the log proves nothing.

For opting out, Anthropic says IP blocking may not work reliably and points to robots.txt. Ask the security team to confirm that verified requests from the intended bots get through, without removing protection.

The content question is whether a reader, or a system summarizing the page, can make a decision from the page alone. Google lists important content available in textual form and structured data that matches the visible text among the fundamentals for its AI features.

Are conditions written down, and the same everywhere?

Price, scope, duration and exclusions should not live only inside an image or behind an interaction. Consider a hypothetical scheduling tool that advertises team booking: a useful page states the supported calendars, account requirements and limits. If the product page, a support article and a comparison post disagree, an AI answer can repeat any of them. Fix the conflict at the source with the product owner's sign-off, and include old pages about retired versions in the check.

Sourced claims, visible authors and consistent brand facts

Attach a source, period and scope to every number, and do not present an experimental result as an expected uplift for every site. Show who wrote or reviewed the page, and remove credentials that cannot be verified.

Use one spelling for brand and product names and add the detail that separates similar products. Google also lists keeping Merchant Center and Business Profile information up to date among its fundamentals for AI features. On brand mentions, Google's guide says that seeking inauthentic mentions across the web isn't as helpful as it might seem.

Google's AI optimization guide has a mythbusting section listing work you can ignore for Google Search. It includes chunking content into tiny pieces, rewriting content just for AI and seeking inauthentic mentions, and says there is no ideal page length. Google also warns that no third-party tool has access to its internal ranking or AI systems. Two items need a closer look.

Do you need llms.txt?

Not for Google Search. Google says Search does not use llms.txt or similar AI files. According to Google, creating them will neither harm nor help your visibility or rankings in Google Search.

It also says maintaining such a file for other services that use it is fine. Google added the note on June 15, 2026.

The OpenAI, Anthropic and Perplexity crawler pages reviewed for this article say nothing about their bots reading llms.txt. If you test one, name the system and the result you will observe, and put the test behind access problems and wrong product facts.

Is there special schema for AI search?

No. Google says structured data isn't required for generative AI search and there's no special schema.org markup you need to add. At the same time, it recommends structured data for rich result eligibility.

FAQ markup deserves a separate note. According to the Search Central updates log, the FAQ rich result stopped appearing in Google Search from May 7, 2026. Its documentation was removed on June 15, 2026. The How-to rich result was removed in September 2023.

Guides that tell you to add FAQPage markup to every page for AI visibility do not reflect those changes. If your pages carry markup, check that it matches the visible text.

Measure GEO in three layers. The first is impressions in Google's Generative AI performance report, and the second is citations in Bing AI Performance. The third is your own fixed prompt set alongside site visits.

Their definitions differ, so never add them into one total. The Dashboard Builder checks which metrics from different platforms can be compared safely.

Google's Generative AI performance report

Google's Generative AI performance report has been rolled out to all websites as of August 31, 2026. It counts how many times links to your site were shown in AI Overviews and AI Mode. The data breaks down by page, country, date and device.

Discover has its own report. In the chart, two links from your site in the same feature count as one impression.

If the report is missing, the usual reasons are too few impressions or an exclusion in the Search generative AI control. Traffic from these features is also included in the standard Performance report under the "Web" search type.

Bing AI Performance

Bing AI Performance launched in Bing Webmaster Tools as a public preview on February 10, 2026. It covers citations across Microsoft Copilot, AI-generated summaries in Bing and select partner integrations. Metrics include total citations, average cited pages per day, page-level citation activity and grounding queries.

Microsoft notes that grounding queries are a sample of citation activity. It adds that citation counts do not indicate placement or ranking inside an answer. The same post recommends IndexNow for notifying search engines about content changes.

Fixed prompts, visits and conversions

For ChatGPT, Claude and Perplexity, rerun the fixed prompt set under the same conditions and compare it with the baseline. On the site, track AI referral visits and the actions that matter. Google suggests tools such as Google Analytics for conversions and time on site.

Note visits without referral data as a limitation. Calculate conversion rates only for the visit group you defined. The denominator rules in the conversion rate optimization guide apply unchanged. Seasonality, campaigns and platform updates also move results, so a before-and-after difference alone does not prove cause.

Fix findings that block access to priority pages or give customers wrong decision facts first. Settings that conflict with your goals come next, then sourcing and content gaps. A template problem that affects many URLs goes ahead of a single-page issue, because one fix closes many findings.

The example table sets no universal time limits. It shows what evidence closes each item.

The first four rows affect a person or a bot directly, so they outrank content improvements. Give each task one owner: the technical team closes access, the product owner closes commercial facts and the editor closes explanations. A ticket marked done without the tested URL and result leaves the next reviewer guessing.

How often should a GEO audit be repeated?

Repeat it on events and on a schedule that matches how fast your content changes. Reopen the affected items when a price, scope, service area or page template changes.

After a robots.txt or CDN change, check again once the roughly 24 hours that OpenAI and Perplexity mention have passed. After changing the Search generative AI control, allow the few days Google describes.

Rerun the fixed prompt set under the same conditions each round. Add new records with dates rather than overwriting old ones, so you can see when a wrong fact appeared and which fix removed it.

A typo fix does not call for a full audit. Close what has evidence and keep uncertain results open, with their limits written down.