Why AI Engines Can Find Your Brand Name but Not Your Category

Short version. Every night from September 11 to September 19 we asked the OpenAI stack nine questions a B2B founder asks before buying help with AI search visibility. No brand names in the prompt, three fresh sessions per question, 243 answers over nine nights. ClawWorld appeared in none of them. On the same nights, questions that include our name came back describing the right company in 5 of 5 answers, most with a citation to our site. In August those same questions returned a claw-machine arcade.

So the name problem is fixed and the category problem is not started. They look like one problem from the outside ("AI doesn't know us") and they are two, with different causes and different fixes. This post is about the difference, and about what we changed on our own site this week, because that part you can copy.

The name problem is an evidence problem

When you ask an engine about a company by name, it resolves the name against whatever public evidence exists: the site, a LinkedIn page, a directory listing, a Crunchbase profile, old news. If that evidence is thin or inconsistent, the engine either says it does not know you or picks the strongest candidate with that name. For us, in August, that candidate was an arcade in Las Vegas. Its LinkedIn page was cited 41 times in our measurement; ours, 12.

The fix is consistency, not volume. One sentence that says what you are, repeated byte for byte everywhere an engine reads: the home page, the structured data, LinkedIn, Crunchbase, llms.txt. Ours is 144 characters and it says what we are not: "ClawWorld is an AEO (Answer Engine Optimization) service for B2B startups based in New York; it is not the claw-machine arcade of the same name." We shipped that in August and again, shorter, on September 22. The identity rate on the OpenAI stack went from 0 of 8 answers to 6 of 8 within a week of the first round of fixes, and it has held.

That job takes weeks, and it is mostly under your control.

The category problem is a retrieval problem

When the question has no brand in it ("best tools for X", "how do I get my SaaS mentioned in ChatGPT"), the engine does not resolve anything. It searches, fetches a handful of pages, and builds the answer from what those pages say. If you are not on those pages, you are not in the answer, however good your own site is.

This month we ran the same kind of test for 50 seed-stage B2B companies, cold, 20 buyer questions each, once on Claude and once on OpenAI's GPT-6 Sol, both with web search on. Some numbers from those 50 reports:

• In the 1,691 answers to questions that did not contain the company's name, the company was named 34 times. Forty of the fifty companies were never named once.
• Those answers carried 12,573 citations. 94 of them pointed at the company's own domain.
• The pages that were cited instead: vendor documentation, comparison posts written by competitors, GitHub repositories, government and standards pages, press, and the companies' own Y Combinator profiles (cited 50 times across the set, more than any of their homepages).

Your own site is a necessary condition. It is not the lever. The lever is the third-party pages the engine already trusts for your category, and the first useful thing a measurement tells you is which pages those are.

What our own site had, and did not have

Before touching anything we audited our own site the way we audit a client's, on September 20. Ten category questions, checked against the raw HTML an AI crawler sees:

• Six of the ten had an answer on the site in a sentence short enough to be lifted. Claude's retrieval quotes at most about 150 characters of a page at a time; a 237-character sentence, however true, does not make it into the answer.
• Three had an answer whose first sentence was over 200 characters. Same problem.
• One question used a word that did not appear anywhere on the site: "GEO", the other name for what we do.
• The Organization schema pointed at a Wikidata item that Wikidata had deleted six days earlier. A dead identity link is worse than none: it tells the engine to merge you with something that does not exist.
• Six public pages had no structured data at all. The blog's article schema had an author and no publisher, so 163 posts belonged to a person named Sammy and not to the company.
• llms.txt linked to anchors on the home page instead of the standalone pages, and described the blog as something it was not.
• Seven URLs from the retired product returned 200 with an empty shell. Bing's index of us had ten URLs; six were those.

None of this is dramatic. All of it is the kind of thing that keeps the name problem from being fully fixed and gives the category problem nowhere to land.

Three things we fixed this week

1. A facts page, and a fuller company record. There is now a page at claw-world.app/facts that states every checkable fact about the company in short rows: what it is, what it is not, where it is, what it sells, what it costs, how it measures, and a dated timeline. The Organization schema on every page now carries an email, a logo, a city, and a mainEntityOfPage pointing at that page. Blog articles carry a publisher. The six pages without structured data have it. The point is one answer, everywhere, in a form an engine can quote.

2. Links that go where they say. llms.txt now points at the pricing, comparison, lab, methodology and learn pages themselves, and it says the blog is daily AI notes plus lab disclosures, which is what it is. The sample report is linked from it too, because a company that sells a free report should show one.

3. Retired URLs answer 410. The seven old paths now return "410 Gone" with a noindex header instead of a 200 shell, and we are asking Bing to drop them. This changes nothing a buyer sees. It changes what an index thinks we are.

The September 22 batch, which this builds on: the 144-character identity sentence at the top of the home page and llms.txt; first sentences under 150 characters for the three long FAQ answers; a sentence that says AEO is also called GEO; the dead Wikidata link removed; robots.txt naming Claude's search crawler.

What we expect, honestly

Not much from this alone, and we would rather say so now than explain it later. The site changes make us quotable. They do not put us on the pages the engines cite for our category, and those are pages we do not own. That is the work of the next six weeks and it happens off-site.

We re-run the same 20 buyer questions at week 4, week 7 and week 8, five samples per engine, with four competitor brands we do no work for measured in the same runs as controls. The nine nightly questions continue. Every round goes on claw-world.app/lab, including the rounds where zero stays zero. The first re-measurement lands in mid-October.

Check your own site in ten minutes

1. Ask the engine your category question with no brand in it, five times in fresh sessions, and write down every cited URL. That list is your actual to-do list.
2. Fetch your most important page the way a crawler does (curl -A "GPTBot" https://your-site/page) and read the first sentence under each heading. If it is not in the raw HTML, or it is over 150 characters, the engine cannot lift it.
3. Ask about your company by name. If the answer describes someone else, do the name work first; category work on top of a wrong identity gets attributed to the wrong company.

Disclosure

ClawWorld is an AEO (Answer Engine Optimization) service for B2B startups — AI agents that get your product mentioned, cited, and recommended by ChatGPT, Perplexity, Gemini, and Google AI Overviews, with a measure–execute–remeasure loop. The measurements in this post are our own nightly scans and the 50 free baseline reports we produced for outreach in September 2026; company names in those reports are not published. The method is written up at claw-world.app/methodology.