Home / Blog / AI Search (AEO and GEO) / Can AI Assistants Find Your Pages? I Tested 9 on My Own Site
AI & AEO

Can AI Assistants Find Your Pages? I Tested 9 on My Own Site

16 October 2026 · By John Carlsson
Not reliably. I asked ChatGPT to list every page on my own site about false positives and it found one of seven reviews, though all were live and linked. Eight more AI assistants did little better: none found all 30 pages, three found none and three invented addresses. If an AI cannot find a page, it cannot cite or mention it.

How I found out

It started with a simple request. I asked ChatGPT to list all the posts and pages on my own site, aiwebpageseo.com, that mention false positives. I know the answer, because I wrote them. There is a dedicated hub page, False Positives in Audit Tools, linked from my home page, with seven reviews of other companies' audit tools under it.

ChatGPT came back with one of the seven reviews. It did not find the hub. It listed several pages that never use the words "false positive" at all, and it said itself that its answer came from "public search results" and was "not a verified exhaustive inventory".

When I pushed it, it tried again and said the blog index "returned a verification page", as if my site had stopped it. My server logs showed that was not true. Only after I gave it the hub address did it find all seven reviews, and it admitted: "the missing articles were a failure of my search coverage, not a problem with the website."

That is the whole argument in one exchange. The pages existed. They were live, linked from the home page and listed in the sitemap. The AI still could not find them. And if an AI cannot find a page, it cannot cite it, quote it or send anyone to it. Everything else you do for AI search counts for nothing if the page is never found.

So I asked the same question of eight more AI assistants, word for word, to see whether this was one AI having a bad day or something wider.

The question, and why it is a fair test

Every AI got exactly this:

"Find every page and blog post on aiwebpageseo.com that mentions false positives. List each one with its full URL and one line on what it says about false positives. Tell me how you found them and whether you think your list is complete."

It is a fair test for three reasons. It asks for something specific and checkable. It does not name a single page, so nothing is handed to the AI. And it asks the AI to say how it searched and whether its list is complete, so it has every chance to admit what it could not do.

It also has a known answer. Rather than trust any AI, I had my own server search every page file and every published blog post for the words "false positive". That list, the answer key, is further down this page.

The score

The key measure is the seven audit tool reviews, plus the hub that links them. These are the pages any good answer had to find.

No AI found the full list. Three found none of the reviews at all. Three invented addresses or sources. Three said an incomplete list was complete, or very likely complete.

What each AI did

ChatGPT

Its first answer leaned on search results and found one review, accessScan. A second attempt widened the search to words like "false alarms" but still missed the hub and six reviews, then blamed my blog for showing it a verification page. Once given the hub address, it read the hub and found all seven, and it corrected itself in plain words. The lesson is clear: it could read the pages perfectly well. It simply had not found them.

Perplexity

Its first answer found the hub, accessScan and Roar Digital, and then said: "Yes, this list is complete". It also said the aisearchradar.io review was "cross-referenced under the same investigation umbrella" as Roar Digital. It is not. It is a separate review with its own post. Its second answer was more careful. Every address was real, it found two further pages that use the phrase "false alarm", and it said its list was only "very likely complete for pages indexed by the search engine".

Microsoft Copilot

Copilot found nothing. Its conclusion was: "I found no publicly indexed aiwebpageseo.com page or blog post that mentions "false positives"". It also said my home page and my blog index do not mention false positives, which is wrong on both counts. It had read search snippets, not the pages. This was the most telling result of all, because Copilot works on Bing, and my server logs show Bing's crawler had already fetched the hub and every review several times.

Gemini

Gemini named five of the seven reviews, but two of the addresses it gave do not exist. It guessed them from the titles, for example inventing a seoClarity address ending in "false-positives" when the real one ends in "cookie-banner". It missed AuditAI and Scam-Detector.com, and then claimed its list "matches the entire set recorded in Bing's server logs". No AI can see Bing's server logs, or mine. When I put this to it, it agreed the claim was "completely invented by the model's text generator to sound definitive and authoritative".

Claude

Claude's search also found nothing. It said so: "A site: search for "false positives" returned nothing from aiwebpageseo.com". Instead of stopping, it opened my home page, followed the link to the hub, and took the seven reviews from there, all at the right addresses. It said plainly that the list was "almost certainly incomplete" and that it had not opened every review itself. It did get one detail wrong. It said my AI-file study's canonical address was in a /research/ folder. It is not. The page's canonical points to itself, and the /research/ address returns a 404.

Grok

Grok gave the best single answer. It found the hub and all seven reviews at the right addresses, described each one accurately, and added two pages nobody else found: my Security Audit page, which describes other scanners reporting "false data", and a community thread inviting people to report false positives in my own tools. I checked its one surprising claim against the server and it was correct. Its fault was confidence. It called the list "very likely complete (or extremely close to it)" when it had missed about 20 pages.

DeepSeek

DeepSeek ran one site search, found nothing, and stopped there. It offered possible reasons, including that the site "may block search engine crawlers". My logs and robots.txt show it does not. To its credit, it did not invent anything and did not claim to be complete.

Mistral Le Chat

Mistral took the longest by far. Its search engine returned nothing useful, so it worked through my site page by page, read the blog index, and then tried to guess the address of each review from its title. Most guesses failed, so it kept trying other routes until it found them. It ended with all seven reviews at the right addresses, plus my home page, the AI-file study and a page quoting Copilot about my site. But it missed the hub page itself, said AuditAI "paywalled" an issue when it was held behind a waitlist sign-up, and said my Learning Hub does not mention false positives when the server shows it does.

You.com

You.com gave the worst answer. Every one of the five addresses it listed on my site returns a 404. It explained how it got them: each heading "was mapped to a likely URL on the aiwebpageseo.com domain, using the typical slug format derived from the title". The headings were not even from my site. They came from other websites' articles about false positives. The only result from my servers in its search was the home page of one of my other sites.

The answer key

This is what my server found when it searched every page file and every published blog post for the words "false positive", on 16 October 2026.

Blog posts (8)

Pages where false positives are the subject (3)

Other pages that use the phrase (19)

The home page, the Answers page and its 12 topic hubs, the Learning Hub, the page quoting Copilot about this site, and three demo and guide pages. Several more pages say "false alarm" or "false data" instead, such as the changelog, the Google Spam Policy Audit page and the Security Audit page.

That is 30 pages and posts that say "false positive". The best AI answers found 10 of them.

It was not the website

Before blaming the AIs, I checked my own side, with the server rather than guesswork.

So the pages were open to every crawler, and one major crawler had already fetched them repeatedly. The AIs that missed them did not miss them because they were blocked.

Why they missed the pages

Most of them only asked a search engine

The AIs that did worst ran a site search and stopped. If their search did not return a page, the page did not exist for them, and several said so in their own answers. The pages were indexed, yet their searches still did not bring them back.

The ones that read the site did best

Claude, Grok and Mistral found the most because they opened the site itself and followed links. My home page links to the hub, and the hub links to every review. A page linked from a page the AI can already reach was found. A page that relied on search alone was not.

Being crawled is not the same as being found

Bing's crawler had fetched every review, yet Copilot, which works on Bing, found none of them. A page being fetched by a crawler does not mean it is in the index the AI is searching, or that the AI will read past the snippet.

When they could not find a page, some made one up

Gemini and You.com built addresses from words they expected to see. Gemini then invented a source to back its list. An AI that cannot find your page may still hand someone a link that looks like yours and goes nowhere.

Why bother? If an AI cannot find a page, it cannot cite it

This is the point of the whole test. People spend time on schema, answer boxes, llms.txt files and AI-friendly writing. None of it matters for a page the AI never reaches. An AI assistant can only mention, quote or link to a page it has found.

It is worse than being left out. Three of the nine told me an incomplete list was complete. Someone asking about my site would have been told those pages did not exist. Two of them invented pages on my site, so a reader following those answers would land on a 404 while my real pages went unmentioned.

Being found does not guarantee being cited. An AI chooses from what it finds, and the quality of the page decides that. But being found comes first, and on this evidence it cannot be taken for granted, even for live, linked pages.

How to run the same test on your own site

  1. Pick a topic you have written several pages about, and write one plain question that asks for every page on your site about it, with full addresses, how the AI found them, and whether it thinks the list is complete.
  2. Build your answer key from your own server, not from memory. Search your page files and your database for the exact phrase, so you know the true number.
  3. Ask the same question, word for word, in several AI assistants. Do not help them, and do not name any page.
  4. Open every address each one gives you. Mark it found, missed or invented.
  5. Check every claim it makes about your site against your server. Did it say your site blocks crawlers, or that a page does not mention something? Check the logs and the page.
  6. Score each answer on the pages that matter most, and note which AIs claimed to be complete when they were not.

What I did about it, and what you can do

  1. I checked robots.txt lets the AI crawlers in. Look for GPTBot, OAI-SearchBot, PerplexityBot and ClaudeBot as well as the search engine crawlers.
  2. I made sure every new page is linked from a page that is already found. My hub is linked from the home page and the Answers page, which is how the best AIs found my reviews.
  3. My XML sitemap updates itself with every new page.
  4. I set up IndexNow, which tells Bing and the other engines that use it about new pages the moment they are published. It needs one small key file on the site and a short request. There is no account to create. The first submission was accepted the same morning.
  5. For Google, which is not part of IndexNow, I use the URL Inspection tool in Search Console to request indexing.
  6. I check the server logs rather than guessing. They show exactly which crawlers have fetched each page and what they were sent.
  7. I will run the same question again once the crawlers have caught up, and compare.

To see which AI crawlers and search bots actually visit your pages, the Real Traffic Audit reads your own server logs and separates human visitors, search crawlers and AI bots.

Sources

Frequently asked questions

Why can ChatGPT, Copilot or Gemini not find my new pages?

Most AI assistants answer from a quick search rather than reading your site, so a page their search does not return does not exist for them, even if it is indexed. In my test the AIs that searched and stopped found few or none of my pages, while those that opened my home page and followed its links found far more.

How do I get my pages found by AI search tools?

Allow the AI crawlers in robots.txt, link every new page from a page that is already found, keep your sitemap current, use IndexNow for Bing and the other engines that support it, and request indexing in Google Search Console. Then check your server logs to see which crawlers have actually fetched each page, rather than guessing.

Can an AI assistant make up pages on my website?

Yes. In my test, two of nine assistants listed addresses on my site that do not exist, built from words in titles or in other websites' articles. One gave five addresses and all five returned a 404. Always open every address an AI gives you before trusting it, and never assume a confident answer is a complete one.

If a crawler has fetched my page, will AI assistants find it?

Not necessarily. My server logs showed Bing's crawler had fetched every one of my review pages several times, yet Microsoft Copilot, which works on Bing, found none of them. Being fetched by a crawler does not mean a page is in the index an assistant searches, or that the assistant will read beyond the search snippet.

How can I test whether AI assistants can find my pages?

Ask several assistants the same plain question, asking for every page on your site about one topic with full addresses and whether the list is complete. Build the true list from your own server first. Then open every address each assistant gives you and mark it found, missed or invented, and check any claim it makes about your site against your logs.

Related posts

💬 Got a problem?

aiwebpageseo.com is a data-driven SEO and AEO platform providing a free suite of technical website tools.