Why Doesn't ChatGPT Know About My Business?

Almost always one of three things, and they are not equally likely.

Either an assistant physically cannot fetch your website, or it can fetch it but was never looking at business websites for that kind of question, or it read your pages and found nothing specific enough to repeat out loud. The first is the most common and by far the least visible — your site loads perfectly in a browser, your robots.txt permits everything, and the request is still being turned away by something you did not know was running.

Here is how to tell which one you have.

Start by dropping the training assumption

Most owners picture the model as having learned about their business at some point, the way it learned about Napoleon. From there the conclusion is obvious and wrong: we are too small, too new, too local to have made the cut, so there is nothing to be done.

That is not the mechanism. When someone types who does composite decking in Bellevue, the assistant runs a search, fetches pages, and writes an answer from what comes back — in the seconds after the question is asked. Almost nothing about your business needs to live inside the model. It needs to be retrievable.

Which is better news than it sounds, because a pipeline can be inspected. Being unknown is not a judgement on your size. It is a description of a break at one of three points.

One: it cannot reach you

This is the failure that costs the most and shows the least, so start here even though it feels too technical to be the answer.

OpenAI does not have "a crawler". It has three, doing three different jobs:

Plenty of businesses blocked GPTBot deliberately, on the reasonable view that their content should not train someone else's model for free. That decision costs nothing in visibility. The expensive version is the one made in a hurry — blocking everything with GPT or AI in the name, or accepting a plugin's "block AI crawlers" default, and taking OAI-SearchBot down with it.

Then there is the failure underneath that one.

Your robots.txt is a request. Your bot protection is a decision. When they disagree, the bot protection wins, and nobody sends you an email about it.

Bot protection, attack-challenge modes, and aggressive rate limits sit in front of your site and make their own judgement about who is asking. A crawler arriving from an unfamiliar data centre at an unusual rate looks exactly like the thing those systems exist to stop. It gets a 429, or a JavaScript challenge page it cannot solve, and it leaves. Your robots.txt said yes. The edge said no. The crawler recorded a failure and moved on, and the only place that event is written down is a log file you have probably never opened.

This is not a rare edge case — it is now one of the most common reasons a site is absent from AI answers while its owner is doing everything the guides say. It also has an ugly property: it usually gets switched on for a good reason. Someone had a genuine scraping problem, or a hosting bill that spiked from bot traffic, turned on the strictest available protection, and solved that problem completely. The cost landed somewhere nobody was watching.

The check: open yoursite.com/robots.txt and read it. Look for the five names — GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, PerplexityBot. If a bot finds a group matching its own name it ignores the general * group entirely, so a permissive * rule does not save you from a restrictive named one. Then find out whether your host, CDN, or security plugin has bot filtering or an attack mode enabled. That switch is not in your website, which is exactly why it survives redesigns.

Two: it was not asking you

The second reason is not a fault in your site at all, and it reframes what your site is for.

In July 2026, a study by Acromatico ran 100 live searches — 10 service verticals across 25 US cities — asking Gemini 2.5 Flash with live search grounding a single repeated question: what are the best [vertical] in [city]? List specific businesses and cite your sources.

Directory and ranking sites were cited in 78 of the 100 answers. The five most-cited domains were bestlawfirms.com, reddit.com, forbes.com, superlawyers.com and justia.com. Business websites appeared constantly — and essentially never as a leading source. The single most-cited domain, bestlawfirms.com, turned up in 16 answers; no individual business site came close.

Take that finding for exactly what it measures: one model, one grounding setup, and a discovery question from someone who does not yet know your name. Ask a different question — is Harbor Deck & Fence any good, does anyone in Bellevue do cedar — and the shape of the answer changes, because now your own pages are directly relevant. But the discovery question is the valuable one. It is the moment a customer with no shortlist gets handed one.

So the honest conclusion is uncomfortable and useful:

A website rarely gets you named. It gets you corroborated.

The assistant finds candidates where it has learned candidates live — directories, industry rankings, review platforms, genuine discussion threads. It then checks those names against whatever else it can pull, and your site is the highest-authority thing available for that check. If the two agree, you survive into the answer with a confident sentence attached. If your site is vague, contradicts your listings, or cannot be fetched, you are the safest name to drop.

Which is why "we'll fix this by adding schema to the website" is such a persistent piece of bad advice. It is work on the corroboration step by someone who has not checked whether the discovery step ever reaches you. The same applies to adding an llms.txt file — useful in its place, and its place is after all of this.

Three: it read you and found nothing to repeat

If a crawler reaches your pages and you still go unnamed, the problem is that an assistant cannot responsibly say anything specific about you.

Assistants are built to avoid making claims they cannot support. Faced with a page that says quality craftsmanship, serving the greater metro area since 2009, we pride ourselves on customer satisfaction, there is no sentence to construct. Every business says that. There is nothing to attach to a recommendation, so you become at best a name in a list, and more often nothing at all.

What has to be findable, in a form a machine can lift:

• The exact service, in the words customers use. Not outdoor living solutions. Composite decking. Cedar fencing.
• Where you actually work. Named towns, not "the greater area". A service radius is a real answer; a vague gesture is not.
• Proof sitting next to the claim, not on a different page. Fourteen projects with photographs, the permit type you pull, the certification and who issued it. An assistant will quote a specific verifiable fact and will skip an adjective.
• The answers to the questions people actually ask before buying — cost range, timeline, permits, materials, what happens when it rains. Those questions are the queries. A page that answers one is a page that can be retrieved for it.
• A page that can be read without running JavaScript, because several of these crawlers do not.

None of that is decoration. It is the difference between being mentioned and being recommended.

A twenty-minute self-check

Do these in order. Stop at the first one that fails, because the ones below it will not matter until it is fixed.

1. Open yoursite.com/robots.txt. Are OAI-SearchBot, ChatGPT-User, ClaudeBot and PerplexityBot permitted? Named groups override *.
2. Find out what sits in front of your site. Ask whoever manages your hosting whether bot protection, an attack mode, or rate limiting is on. If nobody knows, that is your answer to chase.
3. View source on your main service page and search the raw HTML for your headline service. If the text is not there, it was drawn by JavaScript, and some crawlers never saw it.
4. Ask an assistant, with search enabled, what your business does — by exact name and city. Read which sources it cites. If it cites your own site, retrieval works and your problem is further down. If it cites nothing, go back to step 1.
5. Ask it the discovery question your customer would ask — best [your trade] in [your city] — and note every source it names. That list is not competitive research. It is the map of where your category is actually decided.
6. Open your three biggest listings and check the service names, categories and service area against your website, word for word. Assistants cross-check. Disagreement reads as uncertainty, and uncertainty reads as risk.

What you can fix this week, and what you cannot

Some of this is genuinely a Tuesday afternoon.

Correcting robots.txt, making listing names and categories match your site exactly, putting real town names on your service pages, and moving your proof up next to the claim it supports — none of that needs a developer, and all of it helps.

Two things do not go that way.

Bot protection is a balance, not a switch. Simply turning it off exposes you to the scraping and bandwidth costs it was blocking — and on usage-based hosting that arrives as a bill. The work is to let the crawlers that bring you customers through while continuing to stop the rest, which means knowing which agents to verify, how to verify them (published IP ranges and reverse DNS, not the user-agent string, which anyone can type), and where in your stack the rule belongs. Getting this wrong in the cautious direction is what put you here in the first place.

Discovery presence is a campaign, not an edit. Finding out which platforms decide your category in your city, and earning a place on them, takes research and time. It is different work from anything that happens on your website, and no amount of on-site optimisation substitutes for it.

The pattern worth remembering: the parts you can do yourself are the parts that stop you being disqualified. The parts that get you chosen take longer.

What this is called, for anyone searching the term

You will see GEO (generative engine optimisation) and AEO (answer engine optimisation) used for this work, more or less interchangeably, along with AI search optimisation. The vocabulary has not settled and it does not matter much.

What matters is that the object of the work changed. Classic SEO competes for a ranked position on a page of links. This competes to be the business an assistant is willing to name, in a single answer, with a specific reason attached — which rewards being verifiable far more than it rewards being optimised.

If you want a second opinion before spending anything, our free AI Search Snapshot returns one or two findings tied to real pages on your site — including whether crawlers are actually being let through. If you would rather see the whole picture first, the Upgrade Blueprint is $299 and comes off the cost of the work if you decide to go ahead.

And if your site is older than the problem — if it was built before any of this existed and cannot really be fixed in place — that is a different conversation, and an honest one to have early.

Being found is only half of it. What a visitor decides in the eight seconds after they arrive is the other half, and it is a separate problem with separate fixes.

Sources

• Acromatico — The 2026 AI Recommendation Study. 100 live searches across 10 service verticals and 25 US cities, run against Gemini 2.5 Flash with Google Search grounding, July 2026. Source of the 78%-of-answers figure, the top-cited domains, and the finding that business websites rarely rank as a leading source.
• OpenAI — Overview of OpenAI crawlers. The official description of GPTBot, OAI-SearchBot and ChatGPT-User as separate agents with separate jobs.
• The pattern of crawlers being blocked at the CDN or WAF layer despite a permissive robots.txt is drawn from bot-behaviour and server-log reporting across the search-industry press rather than from a single study, and matches what shows up in practice when a site's edge protection is checked directly.

Figures checked 2026-07-31.