AI CMOStudio
Blog

AI search

Why doesn't ChatGPT mention your website?

Because your homepage is not where the answer comes from. When someone asks ChatGPT, Perplexity or Gemini which product to use, the answer is assembled from pages all over the web that discuss businesses like yours, and from whatever the assistant can fetch in the moment. If nothing outside your own site says plainly what you do and who it is for, there is nothing to quote. Here is the five-minute check, and what to fix first.

Yana · Founder, AI CMO Studio23 September 20267 min read
COMPARISON POST“We tried six”COMMUNITY THREAD“Which one for me?”DIRECTORY ENTRYCategory, price, cityREVIEW PAGERating and a quoteTHE QUESTION“best scheduling appfor dog groomers”THE ASSISTANTassemblesfrom what it can readTHE ANSWER“Three worth a look:A, B and C.”2–3 NAMES. YOURS OR NOT.YOUR HOMEPAGEone input among many
One question, one answer, and the pages it was assembled from. Your homepage is in there — as one input, and the one every assistant discounts, because every homepage says it is the best.

I ran our own check on aicmostudio.ai on 9 September 2026. The site scored 100 out of 100 on the machine-readable half: an llms.txt at the root, schema on the pages, robots rules that let the AI crawlers in. It was named in 0 of the 10 answers that came back. What did come back: bigger tools from the categories either side of mine, and two sites whose whole job is writing about tools like mine.

A perfect file score and no mentions. That gap is what this post is about, because it is the normal case, and because the half everyone starts with is the half that decides the least.

What “mentioned” means

When someone asks ChatGPT, Perplexity or Gemini a buyer question in your category, the answer either names your business or it does not. There is no page two and no “you came fourth”. One answer, two or three names in it, and you are in it or you are not.

Four outcomes are worth telling apart, because each one has a different fix.

Cited
The answer links a page on your site. The assistant read you and passed the reader on. Keep that page current.
Named, not linked
The answer says your brand and sends the reader nowhere. It knows you exist; it had nothing of yours worth linking.
Not mentioned
The answer names other businesses. The common case, and the one this post is for.
Doesn't know the name
You ask about yourself by name and the assistant has nothing, or confuses you with someone else. The floor.

Why assistants skip most sites

The answer is assembled from other people's pages. A recommendation question is answered from what the web says about a category: comparison posts, a thread where someone asked the same thing, directory entries, reviews, a newsletter that mentioned you once. Your homepage is one input among those, and the weakest kind, because everyone's homepage claims the same things. If no page outside your own domain puts your name next to the problem you solve, there is nothing to assemble you from.

The fetch may never happen. When ChatGPT goes and looks at a page mid-answer, it arrives as a named agent. OpenAI documents four: OAI-SearchBot, used to surface sites in ChatGPT's search features; ChatGPT-User, for pages visited because a user asked; GPTBot, for model training; and OAI-AdsBot. Perplexity, Google and Anthropic publish their own. A robots.txt that blocks them, or a plugin that helpfully blocked “AI scrapers” for you last year, takes you out of the half of this you actually control.

Nothing on the page is quotable. An answer needs a sentence it can lift. “Trusted by innovative teams” cannot be lifted. “Scheduling software for dog groomers, $19 a month, UK and Ireland” can. Most homepages are written to impress a visitor who already arrived, not to be quoted to someone who has never heard of you.

Check your own site in five minutes

By hand, with no account and no tool. Do it before you change anything, so you have something to compare against.

  1. Minute 1

    Write three questions a buyer types before they know your name

    Not your brand name — the question that comes earlier. Three shapes cover most of it: best <category> for <who you serve>, <category> alternatives to <the obvious leader>, how do I <the problem you solve>.

  2. Minutes 2–3

    Ask all three assistants, in a fresh chat

    ChatGPT, Perplexity, Gemini, one question each, in a temporary or logged-out chat — otherwise you get the answer your own history has trained, not the one a stranger gets. Write down every business named, in the order it appears.

  3. Minute 4

    Open yoursite.com/robots.txt

    Search it for OAI-SearchBot, ChatGPT-User, PerplexityBot, Google-Extended, ClaudeBot. A Disallow next to any of them is a decision somebody made. If it was not you, change it.

  4. Minute 5

    Open yoursite.com/llms.txt

    A 404 means there is no plain summary of your business written for a machine to read. The file is a proposal by Jeremy Howard from September 2024: one markdown page at the root, what you are and where to look, in order. Nobody guarantees an assistant reads it. It costs an afternoon at most.

FIVE BUYER QUESTIONSCHATGPTPERPLEXITYGEMINIbest <category> for <who you serve><category> alternatives to <leader>how do I <the problem you solve><category> for <use case>is <your brand> any good?CITED YOUR SITENAMED YOUR BRANDNOT MENTIONEDDOESN’T KNOW IT
Five questions, three assistants, fifteen cells — one mark each. Example layout, not a real site: what you are reading is which questions you lose and to whom, not a score.

What the result tells you, and what it doesn't

Read it question by question. Each of the four marks points at a different job:

  • Cited — that page works. Note what makes it quotable and do the same on the next one.
  • Named but not linked — the model knows the name from somewhere old. Give it something current to link: one page that answers that exact question.
  • Not mentioned — read who was. Those names are your real competitive set in this channel, whatever your usual competitor list says. Ask the assistant which sources it used, then go and read them.
  • Doesn't know you — nothing outside your domain names you. One honest mention, anywhere a person would actually look, is worth more here than a month of homepage edits.

What it will not tell you is why. The assistant does not publish its reasoning, so the result is an outcome, not an explanation. And one run is a reading, not a verdict: answers move with the wording of the question and with every model update, and two runs an hour apart can disagree. That is true for everyone in your category as well, which is why the useful habit is running it again after you change something — not every Monday.

Three fixes, in the order you control them

TIME UNTIL IT CAN SHOW UP IN AN ANSWER01A page worth quotingYou publish itLIVE ON DEPLOY02robots.txt, llms.txt, sitemapYou publish itLIVE ON DEPLOY03A mention on someone else’s pageThey publish itWEEKS, IF IT LANDS
Two of the three are yours to publish and live the moment you deploy them. The third is the one that moves the answers, and someone else has to press the button.

1. Make one page quotable

A facts page, or the top of your homepage: one sentence per fact, no adjectives. What it is. Who it is for. What it costs. Where it works. What it does not do — that last one gets quoted more often than you would expect, because people ask assistants “which is right for me” far more often than “which is best”, and a limit is the easiest thing in the world to quote.

2. Hand the machines a map

A robots.txt that names the agents and allows them, an llms.txt at the root, a sitemap that is actually current. Ours are public: aicmostudio.ai/robots.txt and aicmostudio.ai/llms.txt. Copy the shape.

llms.txt

# AI CMO Studio

> AI CMO Studio is an AI marketing studio for founders, solo
> businesses and small teams who run one or several websites.

Site: https://aicmostudio.ai
Contact: support@aicmostudio.ai

## What it does

- Research: reads up to twelve pages of your website and writes
  seven working documents from it …
The first lines of our own llms.txt. Name, one paragraph of what it is, then the links, in the order a stranger would need them.

The free llms.txt generator writes the file for you: it reads up to eleven pages of your site and returns a summary, your key pages with a line each, and your key facts — built only from what your pages actually say. Anything they do not state is left out rather than guessed, which is the point: a file full of invented claims is worse than no file.

3. Get named on a page that isn't yours

This is the one that moves the answers, and the slow one. Start from the sources the assistant already used for the question you lost — it will list them if you ask. Then: the roundup or directory your category actually has; the community thread where your buyer's question is asked every week, answered properly, with your own product mentioned once at the end along with what it is not for; a comparison written by someone who has used both.

What does not work: paid listings on pages nobody reads, and dropping your link into threads without answering the question. Both are visible to a reader, and a page nobody quotes is a page no assistant quotes either.

Fixes 1 and 2 are live the moment you deploy them. Fix 3 has no timeline anyone can honestly promise you: someone else has to publish it, and the assistants have to find it. Re-run the check when each one lands — that is what the saved link is for.

Next step

See which questions your site loses, and to whom.

Run the AI Visibility Check

Free · no account · about 30 seconds

Written by YanaFounder, AI CMO Studio. She runs the studio on her own websites first.
Last updated: 23 September 2026.

Questions

Asked about this one

Is this just SEO with a new name?

The inputs overlap — crawlable pages, clear language, links from places that matter — but the output is different. Search gives a reader ten links to choose from; an assistant gives one answer with two or three names in it. Ranking eleventh is a bad day in search and nothing at all here.

Should I block GPTBot?

Decide it separately from the rest. Per OpenAI's own documentation, GPTBot is the training crawler, while OAI-SearchBot is what surfaces sites in ChatGPT's search features and ChatGPT-User is what fetches a page because someone asked. Blocking the training crawler is a real choice with a real cost; blocking the other two removes you from the answer itself.

Will an llms.txt file get me mentioned?

No. It is a map, not a vote. No assistant publishes a guarantee that it reads one. It is worth doing because it is cheap, because it forces you to write down what you actually are in plain sentences, and because it removes one excuse from the list.

I run a local service business, not software. Does any of this apply?

The mechanism is identical: the answer is built from pages other people publish about businesses like yours. The pages differ — local roundups, the review profiles people actually read, the town's own listings rather than a software comparison blog. Run the same five minutes with “in <your town>” on the end of each question.

How often should I re-run the check?

After something changes: a page rewritten, a file published, a mention that finally went live. And once when a new name appears in the answers, because that is usually someone else's third fix landing.

The studio behind the tools

This is one afternoon of what your marketing team does every week

AI CMO Studio reads your site once, then its specialists keep working: AI-search fixes, SEO, articles, social posts, newsletters and visuals, every draft in one feed you approve. Built by a founder for her own sites first. Early access opens to the waitlist in order.