The website changes that get a product recommended by AI answer engines are a short list, and on our own site they split into two groups with two different results. The first group makes a page findable and quotable: the words in the raw HTML, one identity sentence repeated everywhere, a first sentence under each heading short enough to lift, structured data that says what the company is, retired pages that say they are gone. The August part of that group, the identity sentence and a content pack with structured data, was followed within a week by the question asked about us by name going from 0 of 8 correct answers to 6 of 8 on the OpenAI stack. The second result is the one most generative engine optimization advice leaves out: after 24 more changes between September 22 and October 10, the rest of that group among them, the same site was recommended for its category in 0 of 200 answers on October 12. The site is necessary. The pages that get a product recommended are mostly pages you do not own.
What an AI engine found on our site before we touched it
On September 20, 2026 we audited our own site against the raw HTML an AI crawler sees, with ten category questions in hand. Seven findings, none dramatic, every one of them something we had told clients to fix:
| Finding | Why it matters to an engine |
|---|---|
| 6 of 10 questions had an answer on the site in a sentence short enough to be lifted; 3 had a first sentence over 200 characters | Claude's retrieval quotes about 150 characters of a page at a time; a true 237-character sentence does not make it into the answer |
| 1 question used a word that appeared nowhere on the site ("GEO") | A page cannot be fetched for a term it never uses |
| The Organization schema pointed at a Wikidata item deleted six days earlier | A dead identity link tells the engine to merge you with something that does not exist |
| 6 public pages had no structured data | Nothing on the page says what kind of thing it is |
| The blog's article schema had an author and no publisher | 163 posts belonged to a person, not to the company |
| llms.txt linked to anchors on the home page and described the blog as something it was not | The file meant to orient a crawler pointed it at the wrong things |
| 7 URLs from a retired product returned 200 with an empty shell | Bing's index of us held 10 URLs, and 6 were those |
The list is the point. Before any of it, in August, questions that used our name came back describing a claw-machine arcade in Las Vegas; its LinkedIn page was cited 41 times in our measurement and ours 12. The engine had thin evidence and picked the stronger candidate with the same name, which is what an engine does.
Generative engine optimization website changes, with the number each one moved
Everything we changed is in our public changelog with a date, so a before-and-after number sits beside the list of what happened between. These are the rows that bear on this question:
| Change | Live | What moved afterwards |
|---|---|---|
| The identity set: one sentence saying what we are, byte for byte the same at the top of the home page, in the structured data and in llms.txt, and an on-site content pack with a FAQ carrying structured data, a comparison page and two explainers | August 2026; the content pack on August 17 | In the round of August 23, by-name answers on the OpenAI stack went from 0 of 8 correct identities in the baseline round of August 16 to 6 of 8, and the same stack cited our own domain for the first time, in 7 of 48 answers |
| A 144-character sentence saying what we are not, the arcade, in front of the identity sentence on the home page and in llms.txt; first sentences under 150 characters for the three long FAQ answers; a sentence saying AEO is also called GEO; the dead Wikidata link removed; robots.txt naming Claude's search crawler | September 22 | Nothing measurable on category questions: 0 of 200 answers on September 28 |
| A facts page with every checkable fact in short rows; Organization schema with an email, a logo, a city and a page it points at; a publisher on every article; structured data on the six pages without it | the week of September 28 | Nothing on category questions: 0 of 200 on October 12 |
| llms.txt pointing at the real pricing, comparison, lab, methodology and learn pages | the week of September 28 | Same |
| The seven retired URLs returning 410 Gone with a noindex header | the week of September 28 | Changes what the index thinks we are; no answer-side number yet |
Read the right-hand column down and the pattern is plain. The identity set moved the identity question, fast. Nine new pages written for the category questions went live after September 26, and in the October 12 round they were cited 0 times across 200 answers. In the same round three vendors we had not seen two weeks earlier entered Claude's answers, each because its own pages were being cited: 12, 12 and 9 times. Their pages were in the index Claude searches. Ours, that week, were not.
Why the site is necessary and still not the lever
A product is recommended when the engine fetches pages for a question with no brand in it and those pages name the product. In September we ran that test for 50 seed-stage B2B companies, 20 buyer questions each, on Claude and GPT-6 Sol with web search on. In the 1,679 answers to questions that named neither the company nor its product, the company was named 81 times, under 5%, and 18 of the 50 were never named once. Those answers carried 12,728 citations; 89 of them, 0.7%, pointed at the company's own domain. The companies' homepages were cited 47 times, their Y Combinator profiles 28. What got cited instead was vendor documentation, comparison posts written by competitors, code repositories, standards pages and press.
The engines also read differently from each other, which changes what "the site" can do. Across all 100 of its answers on October 12, the whole round of 20 questions at five samples each, GPT-6 Sol cited developers.google.com in 72, help.openai.com in 32 and bing.com in 22. On the 12 of those questions that ask how to do something, 60 answers, it named a vendor in 1. Claude, on the question this post answers, searched the web in 4 of its 5 answers on September 28 and in 1 of 5 on October 12, the rest from memory. A page cannot reach an engine that did not search.
The order that follows from our numbers
Do the identity set first, in one week, because it is the only change here with a measured effect and the fix is under your control: one sentence that says what you are, byte for byte the same on the home page, in the structured data, on LinkedIn, in llms.txt, with a FAQ carrying structured data beside it. Then the quotable sentences, under 150 characters, under every heading that answers a buyer question, in the raw HTML. Then check the indexes, because a page absent from Bing or Brave cannot be cited by the engine that reads it however good it is; our own pages were findable by Claude's search on October 15 and gone again on October 19. Then the part that is most of the work: ask your category's questions yourself, five times per engine, and list every page the answers cite. Getting named on those pages is what gets a product recommended, and none of it is a change to your website.
We would rather say the limit now than explain it later. The site changes above make us quotable, and on category questions our own number after all of them is still zero on every round we have published: the three August rounds on our Lab page and the September 28 and October 12 rounds above. That is the honest size of what a website change does on its own.
How we checked
The audit is our own, on September 20, 2026, against the HTML returned to a crawler user agent. The identity and own-domain figures are the OpenAI-stack rows of the August 16 and August 23 rounds on the Lab page. The 0 of 200 figures are the September 28 and October 12 rounds: 20 buyer questions, five samples each, Claude Opus 5 with web search and GPT-6 Sol through Codex with web search, with the two engines' cited domains counted from every answer. The 50-company figures are the free baseline reports we produced in September 2026, recounted on October 4 from each report's own verdicts; the company names are not published. The dates of the changes are the rows of our changelog.
Disclosure
ClawWorld is an AEO (Answer Engine Optimization) service for B2B startups based in New York. We sell the work described here, which is a reason to read the zeros above as carefully as the sixes.