A buyer asks ChatGPT which tool to use, and your brand isn’t in the answer. Or it is, with last year’s price. Either way you lose the deal before the buyer reaches your site, and Search Console shows no sign of it.
Short answerUpdated 13 Oct 2026
A GEO audit checks four things: access, visibility, accuracy and quotability. It shows whether AI tools can reach your pages, what they say about you, and whether they get your facts right. Start with access, because a page crawlers can’t read can’t be cited.
- Check AI crawler rules in robots.txt and in your firewall. Blocking OAI-SearchBot keeps you out of ChatGPT search, while GPTBot only collects training data.
- Make sure page text, prices and structured data sit in the HTML your server sends. Most AI crawlers don’t run JavaScript.
- Check the Search generative AI control in Search Console. Google says a site must be included there to appear in AI Overviews and AI Mode.
- Ask each buyer question more than once. In my test, 5 of the 36 sites ChatGPT cited came back in all five runs of their question.
- Fix facts on your own site first. ChatGPT cited software makers’ own pages in 93 of 106 citations, while Perplexity leaned on roundups.
- Treat structured data as accuracy work and llms.txt as optional. Google says its AI features need no special markup.
A GEO audit finds out why the answer leaves you out. GEO, generative engine optimization, is the work of getting cited, named and recommended in AI answers.
These are the 12 checks I run before any other GEO work. Each one comes from a problem I found in a real audit or test. Each comes with a test you can run yourself.
What is a GEO audit?
A GEO audit is a review of how AI search tools see your brand. It checks whether ChatGPT, Perplexity, Gemini and Google’s AI Overviews can crawl and read your pages. It also checks what they say about you, which sources they trust and whether they get your facts right.
It overlaps with an SEO audit, but it asks different questions.
| SEO audit | GEO audit | |
|---|---|---|
| Main question | Can Google crawl, index and rank this page? | Can AI tools read, quote and recommend us? |
| Crawlers checked | Googlebot and Bingbot | Those two, plus OpenAI’s, Anthropic’s and Perplexity’s |
| Success looks like | Rankings and clicks | Being cited, named and recommended |
| Main evidence | Search Console and a site crawl | Answers to buyer questions, asked more than once |
| Accuracy work | Titles, content and links | Prices and claims everywhere AI tools read them |
An AEO audit and an AI visibility audit are the same job under other names. AEO stands for answer engine optimization. An “AI SEO audit” usually means something else: a standard SEO audit run with AI tools.
The GEO audit checklist: 12 checks at a glance
The checks fall into four groups: access (1 to 4), visibility (5 to 7), accuracy (8 and 9) and quotability (10 to 12). Download the 12 checks as a one-page PDF.
| # | Check | The question it answers |
|---|---|---|
| 1 | AI crawler rules in robots.txt | Do you let AI search crawlers in? |
| 2 | Firewall and CDN blocks | Does anything block them anyway? |
| 3 | Page text in the raw HTML | Can they read the page without JavaScript? |
| 4 | Indexed and eligible for AI answers | Are your pages in Google and Bing, and allowed in AI answers? |
| 5 | A baseline of buyer questions | What do AI answers say about you today? |
| 6 | Each question, more than once | Is that answer stable, or a one-off? |
| 7 | Who gets cited instead | Which sites do AI tools trust in your category? |
| 8 | Your pages agree with each other | Do all your pages give the same facts? |
| 9 | Other sites agree with you | Do profiles and listings say the same? |
| 10 | Answer-first passages | Can a model lift a clean answer from your key pages? |
| 11 | Structured data that tells the truth | Does your markup state true facts in the server HTML? |
| 12 | One clear entity | Do AI tools know which company your name means? |
Which GEO audit tools do you need?
You can run all 12 checks with free tools. You need curl, your server or CDN logs, Search Console, Bing Webmaster Tools, Google’s Rich Results Test and a spreadsheet.
Paid AI visibility tools automate checks 5 to 7 by running your questions on a schedule. They can’t do the checks that need your logs, CDN settings and profiles.
| Checks | Free tool |
|---|---|
| 1 and 2 | robots.txt, curl, and your server or CDN logs |
| 3 | curl, or view source in your browser |
| 4 | URL Inspection in Search Console and Bing Webmaster Tools |
| 5 to 7 | The AI tools themselves, logged out, plus a spreadsheet |
| 8 and 9 | A site: search, and the pages each answer cites |
| 10 to 12 | Your own reading, Google’s Rich Results Test and a brand search |
Access: can AI tools reach and read your pages?
AI tools can only cite pages their crawlers can reach and read. Four checks cover access: robots.txt rules, firewall and CDN blocks, text in the raw HTML, and index and AI eligibility settings.
1. AI crawler rules in robots.txt
Your robots.txt should let each AI company’s search crawler in, and allow or block its training crawler on purpose. OpenAI says ChatGPT search won’t show sites that block OAI-SearchBot, except as navigational links. GPTBot collects training data.2
| Company | Search | Training | Fetches when a user asks |
|---|---|---|---|
| OpenAI | OAI-SearchBot | GPTBot | ChatGPT-User |
| Anthropic | Claude-SearchBot | ClaudeBot | Claude-User |
| Perplexity | PerplexityBot | None, Perplexity says | Perplexity-User |
| Googlebot | Google-Extended (a robots.txt token) | Google-Agent and other user-triggered fetchers |
Fetchers that run when a user asks may skip robots.txt. OpenAI says its rules “may not apply” to ChatGPT-User, and Perplexity and Google say their user fetchers generally ignore them. Anthropic says all three of its bots follow robots.txt.23
Google-Extended controls whether Google uses your pages to train Gemini and to ground answers in Gemini Apps. Google says it doesn’t affect your inclusion or ranking in Google Search.4
Test: open yourdomain.com/robots.txt. Look for User-agent lines that name these crawlers, and for a catch-all Disallow: / that covers them.
Pass: OAI-SearchBot, Claude-SearchBot, PerplexityBot, Googlebot and Bingbot can reach every page you want quoted. You allow or block GPTBot, ClaudeBot, Google-Extended and CCBot, Common Crawl’s bot, on purpose.
A support SaaS I audited let every crawler in, training crawlers included. That passes if someone chose it on purpose. An AI marketing startup allowed more than 20 AI crawlers by name in its robots.txt.5
2. Firewall and CDN blocks
robots.txt is a request. A firewall or CDN rule can turn a crawler away before it reads the file, and robots.txt won’t show it. OpenAI tells site owners to confirm their host or CDN allows traffic from its search crawler’s IP addresses.6
In July 2025, Cloudflare began asking every new domain whether to allow AI crawlers. In July 2026, it announced new defaults for domains added from 15 September 2026. On their pages that show ads, training and agent crawlers would be blocked.7
Cloudflare also said customers who block training will block Googlebot, Applebot and Bingbot, which crawl for both search and training.7
Test: fetch one key page as a browser, as Googlebot and as each AI search crawler. Then compare the status codes and sizes.
for ua in "Mozilla/5.0" "Googlebot/2.1" "OAI-SearchBot/1.0" "Claude-SearchBot" "PerplexityBot/1.0"; do
curl -sL -o /dev/null -w "%{http_code} %{size_download} $ua\n" -A "$ua" https://yourdomain.com/pricing
done
A firewall that checks IP addresses can let this fake user agent through and still block the real bot. Your server or CDN logs settle it: look for requests from these crawlers and the status codes they got. On Cloudflare, AI Crawl Control is available on all plans and shows which AI crawlers reach your site.7
Pass: the same response for every user agent, and logs that show the search crawlers getting 200 responses.
At the support SaaS, nine spoofed AI user agents got the same 200 response as Googlebot, byte for byte.5
3. Page text in the raw HTML
Most AI crawlers read only the HTML your server sends. In a December 2024 study by Vercel and MERJ, the crawlers from OpenAI, Anthropic, Meta and Perplexity didn’t render JavaScript.8 If your prices, FAQ answers or product details load after the page opens, those crawlers miss them.
Test: strip the scripts from the raw HTML, then search for a phrase from the page. Then check that your structured data sits in its own script tag.
curl -sL https://yourdomain.com/pricing | perl -0pe 's/<script\b.*?<\/script>//gis' | grep -c "a short phrase from your page"
curl -sL https://yourdomain.com/pricing | grep -Eic '<script[^>]+application/ld\+json'
Pass: the page text, prices, FAQ answers and structured data all appear in the server HTML.
A satellite operations SaaS kept 5 of its 6 FAQ answers inside a script tag, where crawlers reading page text skip them. Its headlines split words across tags with no spaces, so extracted text ran the words together. At the support SaaS, article markup lived only in the JavaScript payload until the posts moved to server rendering.5
AI app builders often ship pages that fail this check. See what I found on 65 live Lovable sites.
4. Indexed and eligible for AI answers
AI answers mostly cite pages that search engines have indexed. Google says a page must be indexed and eligible for a snippet to appear in AI Overviews or AI Mode. Microsoft says Copilot relies on the same crawling and indexing as Bing search.910
OpenAI says ChatGPT search sometimes partners with other search providers.6 So a page missing from Google or Bing loses that engine’s AI answers, and may lose ChatGPT’s too.
In 2026, Google added a Search generative AI control to Search Console, under Settings. Google says a site must be included there to be eligible for AI Overviews and AI Mode. Including is the default, but a property can inherit Exclude from a parent property.9
Google limits what it shows from a page through nosnippet, data-nosnippet, max-snippet and noindex.9 On Bing, noarchive keeps a page out of Copilot answers.10
Test: run your key pages through URL Inspection in Search Console and in Bing Webmaster Tools, and open the Search generative AI control. Then check the pages for robots meta tags and headers.
curl -sL https://yourdomain.com/pricing | grep -Eio '<meta[^>]*name=.?(robots|googlebot|bingbot)[^>]*>'
curl -sIL https://yourdomain.com/pricing | grep -i x-robots-tag
Pass: your key pages are indexed in Google and Bing, and your site is included in Search generative AI. No page you want quoted carries noindex, nosnippet, noarchive or a small max-snippet.
The AI marketing startup had 8 of its 31 pages in Google and none in Bing. The satellite operations SaaS served its API reference pages with a 404 status before JavaScript drew each page.5
Visibility: what do AI answers say about you now?
Visibility is what AI answers say about you today: whether they cite your pages, name your brand or recommend you. Measure it with fixed buyer questions, asked more than once.
5. A baseline of buyer questions
A GEO baseline is 20 to 30 buyer questions, run on each AI tool and logged the same way each month. Mix “what is [category]”, “best [category] for [use case]”, “[you] vs [rival]”, “[rival] alternatives” and “how much does [you] cost”.
Run each question in ChatGPT, Perplexity, Gemini and Google’s AI Overviews or AI Mode. Use a logged-out or fresh session, so past chats don’t shape the answer. Then place each answer on a ladder with four rungs:
- Read: the tool looked at your page but didn’t link it. It shows in the full source list, or your logs show the fetch.
- Cited: the answer links your page.
- Named: the answer mentions your brand.
- Recommended: the answer tells the buyer to pick you.
| Question | Tool | Run | Read? | Cited? | Named? | Recommended? | Rivals named | Your pages cited |
|---|---|---|---|---|---|---|---|---|
| Best [category] for [use case] | ChatGPT | 1 of 5 | Yes | Yes | Yes | No | Rival A, Rival B | /pricing |
Pass: 20 to 30 questions, each logged per tool and run, with a rung for every answer.
For the satellite operations SaaS, the core buyer question got no mentions on ChatGPT or Perplexity, while ChatGPT named four rivals. For the AI marketing startup, ChatGPT recommended 12 companies for its core use case, none of them the client.5
6. Each question, more than once
AI answers change from run to run, so one answer proves little. On 13 October 2026, I sent six buyer questions five times each to ChatGPT and Perplexity through their APIs, with web search on. None of the questions came from a client’s market.1
Sites cited that came back in all five runs (5 of 36 and 46 of 48)
Brands named in all five runs (7 of 36 and 29 of 38)
Citations that were roundups or reviews (8 of 106 and 275 of 295)
| ChatGPT (API) | Perplexity (API) | |
|---|---|---|
| Answers | 30 | 30 |
| Sites cited | 36 | 48 |
| Sites cited in all five runs | 5 | 46 |
| Brands named | 36 | 38 |
| Brands named in all five runs | 7 | 29 |
| Questions where the top pick changed | 2 of 6 | 2 of 6 |
ChatGPT’s sources and brand lists moved the most. Asked for the best scheduling software for a small physiotherapy clinic, it gave four top picks in five runs. It led with Noterro twice, then ClinicSense, WebPT and Jane.1
Perplexity cited almost the same sites every time, yet its top pick still moved. For the password manager question, three runs led with 1Password, one with Bitwarden and one with Keeper.1
Test: run each question three to five times and log every run. Then roll the rows up into a rate, such as “named in 3 of 5”.
Pass: your baseline reports rates, and you re-run the same questions each month, so you can tell a real shift from run-to-run noise.
7. Who gets cited instead
Different AI tools trust different sources. In my repeat test, 93 of ChatGPT’s 106 citations were software makers’ own product, pricing or help pages. Perplexity leaned the other way: 275 of its 295 citations were roundups and reviews, 190 of them on software companies’ sites.1
For the password manager question, Perplexity cited PCMag, Wired, ZDNet, TechRadar and Security.org in all five runs.1 In my earlier test of 57 ChatGPT citations on B2B buyer questions, review directories and comparison sites supplied 25. That test used an older model, gpt-4o, so check the mix in your own category.
Test: list every source each answer cites, and sort them into your pages, rivals’ pages, roundups, review directories and forums. Then take the ten third-party sites cited most often, and check whether you’re listed on each one, accurately.
Pass: you know the ten third-party sites that feed answers in your category, and you have a plan to appear on each.
Accuracy: do AI answers get your facts right?
AI answers repeat whatever prices and claims they find, on your pages and on other sites. Two checks find the mismatches.
8. Your pages agree with each other
AI tools quote your pages, old ones included. If your pricing page says one price and a blog post says another, either can end up in an answer.
Test: search Google for site:yourdomain.com "[old plan name]" and site:yourdomain.com "[old price]". Then compare your pricing page, help center, blog calls to action, structured data and llms.txt against one master list of prices and plans.
Pass: one price, one plan list and one product description across the site.
The AI marketing startup listed three social profiles in its markup, and all three differed from the ones in its own footer.5 Old prices linger the same way after a pricing change: in blog posts, help articles and calls to action that nobody reopened.
9. Other sites agree with you
Answers also draw on pages you don’t host: review directories, launch platforms, app marketplaces, Crunchbase, LinkedIn and partner pages.
Test: ask ChatGPT and Perplexity “How much does [brand] cost?”, then open every page they cite. Check your main profiles by hand too.
Pass: every profile you control matches your site on name, price, plans and description.
The satellite operations SaaS’s profile on a startup directory gave a different LinkedIn page and contact details from its own markup.5
Quotability: are your pages easy to quote and hard to confuse?
A quotable page answers each question in its first lines and states true facts in its markup. It also uses a name AI tools can’t confuse with another company.
10. Answer-first passages
AI answers lift passages. A passage gets quoted when it answers a question on its own, in the first lines under a heading.
Test: on your pricing, product and comparison pages, read the first two sentences under each heading. Could a stranger quote them alone? Check that headings match buyers’ questions and that comparisons sit in tables with dated numbers.
Pass: every key section opens with an answer that stands alone.
The satellite operations SaaS had about 600 unique words across its whole site and described its product in three one-line blurbs.5
In my AI Overview case study, the most-cited page opens with a definition, then a table.
11. Structured data that tells the truth
Google says its AI features need no special markup: “There’s also no special schema.org structured data that you need to add.”9 Treat structured data as accuracy work, because wrong markup hands AI tools wrong facts.
Test: run key pages through Google’s Rich Results Test. Check that sameAs links open, the logo is at least 112 pixels square, the type fits your business and prices are real.11
Pass: Organization and Product markup sit in the server HTML, and every value is true.
The AI marketing startup’s logo markup pointed at a 96-pixel favicon. One sameAs link returned a 404, and its founding year was three years off.5
The satellite operations SaaS marked itself up as a local business and gave a currency with no price. It carried FAQ markup on four pages, though only one had a FAQ.5 FAQ markup no longer earns rich results: Google stopped showing them on 7 May 2026.12
12. One clear entity
AI tools need to know which company your name means. If another company shares your name, or you spell it two ways, AI answers can mix you up.
Test: search your brand name alone, and ask ChatGPT “What is [brand]?” in a fresh session. Check that your site, markup and profiles spell the name the same way.
Pass: the results and ChatGPT’s answer describe your company, with one spelling across your site, markup and profiles.
The satellite operations SaaS spelled its product name two ways and shared it with several larger companies. The AI marketing startup shared its short name with a better-known company.5
Use the longest unique form of your name everywhere. Match it on Wikidata, Crunchbase and your review profiles, then link each profile from your markup with sameAs.
Do you need an llms.txt file?
No. llms.txt is a proposed file that points language models and AI agents to a site’s key pages. Jeremy Howard published the proposal in September 2024.13
Google says llms.txt files will “neither harm nor help” your visibility, “as Google Search ignores them.”9 OpenAI, Anthropic and Perplexity don’t say their crawlers read it. The file is cheap to add, so I note it in audits and rank it below checks 1 to 12.
What should a GEO audit deliver?
- A crawler access table: each AI crawler, its robots.txt rule, its response code and what your logs show.
- A baseline sheet: every question, tool and run, with read, cited, named and recommended rates.
- A source map: the ten third-party sites cited most in your category, and whether you’re listed accurately on each.
- A fact list: every page or profile with a wrong price, plan or claim, and the right value.
- A fix list in priority order, with the evidence for each item.
What I fix first
- Access blockers (checks 1 to 4). A page crawlers can’t read or index can’t be cited.
- Wrong facts (checks 8 and 9). A wrong price in an AI answer costs deals before a buyer reaches your site.
- Missing pages for the questions where rivals get cited (from checks 5 to 7). Often that’s a comparison or alternatives page.
- Quotability and entity (checks 10 to 12), starting with the pages that sell.
- Re-run the questions each month, three to five runs each. Run the full audit each quarter, and after a redesign, migration or pricing change.
That’s the order I follow in my AEO and GEO services, starting with this audit.
How to measure GEO after the audit
Track the ladder for every question: read, cited, named and recommended, each as a rate across runs. That’s the main score.
In GA4, open the default AI Assistant channel. Google added it in May 2026 for visits from tools such as ChatGPT, Gemini and Copilot.14 For a custom segment, filter Session source with “matches regex” and this pattern, keeping the .* at both ends:
.*(chatgpt\.com|openai|perplexity|gemini\.google\.com|copilot|claude\.ai|deepseek|grok).*
In Search Console, AI Overviews and AI Mode count inside the Web search type. The Generative AI performance report shows their impressions by page.915 Bing Webmaster Tools has an AI Performance report that shows how often Copilot cites your pages.10
Watch branded searches too. A buyer who meets you in an AI answer may search your name next.
GEO audit FAQ
Is an AEO audit or AI visibility audit the same as a GEO audit?
Yes, in practice. All three check whether AI tools can read your pages, what they say about you and whether they get your facts right.
Can ChatGPT do an SEO or GEO audit?
Partly. ChatGPT can review a page and answer your buyer questions once, and one run isn’t a baseline. It can’t crawl your whole site, see your logs or Search Console, or show how other AI tools answer.
Is there a GEO audit tool that runs all 12 checks?
No single tool covers all 12. Paid AI visibility tools automate the buyer-question checks. Crawler access, CDN settings and profile accuracy still need your logs, your settings and a person reading the pages.
How often should I run a GEO audit?
Run the full audit every quarter, and after a redesign, a migration or a pricing change. Re-run the buyer questions every month.
How much does a GEO audit cost?
Run yourself, it costs time rather than money, because every check here uses free tools. The baseline takes longest: 25 questions on four tools, three runs each, is 300 answers to log. If you want me to run it, I quote after a short brief.
Do I need an llms.txt file?
No. Google says Google Search ignores llms.txt files. Fix crawler access, page text and your facts first.
Limits of this guide
The audit examples come from three companies I audited between July and September 2026. They show the kinds of problems I find. Three audits can’t show how common each one is.
The repeat test used six questions, two tools and one day. The ChatGPT and Perplexity apps can answer differently from their APIs, especially for signed-in users. I matched brand names against a list for each question, so a brand outside the list wouldn’t count.1
I sell this audit as a service, and these are the checks I run for clients. The repeat-test data is free to download, one row per answer. You can reuse it under CC BY 4.0: credit Bildad Oyugi and link to this page.
Want the 12 checks run on your site?
I run the full audit, then hand you a fix list in priority order, with the evidence for each item. Send your URL and the questions your buyers ask, and I’ll reply by email with questions or a rough plan.
Send me your URL