More people now ask an AI assistant about a company than read its About page. So we built a free AI visibility check: it asks ChatGPT, Claude and Grok what they know about a website, once from memory and once with web search switched on, the way the real apps answer, and it compares the answers with what the site says about itself. Then we did the obvious thing and ran it on rank.fast.
The short version: on 31 August, two of the three assistants confidently described a company that is not us. On 3 September, after three days of fixes, one of them had corrected itself and one had not. None of the three knew us from memory on either date, which for a twelve-day-old site is exactly right.
What the assistants said on 31 August
The check runs each assistant twice. From memory alone, all three said they had no knowledge of rank.fast. With web search on, they went looking, and this is what they came back with.
None of that is what we do. Its cited sources included a Facebook page for a different business called RankFast, a Glassdoor listing for a third, and a Wikipedia article on fast food chains.
It cited our own homepage as the source for the domain being parked, which means it was working from a copy of the page fetched before we launched, and then substituted a company with a similar name.
Correct in every particular. Grok ran about a dozen searches before answering and cited our day-zero post. The other two stopped early and filled the gap with a neighbour.
Why a correct website was not enough
Our own pages were fine on every mechanical check the tool runs: none of the ten AI crawlers blocked, an llms.txt file, Organization structured data, no JavaScript needed to read the content. The problem was outside the site. Nothing on the web described rank.fast except rank.fast, and the name space is crowded: rank-fast.com sells a website template, rankfast.co is an agency, RankFast LLC is another, and a fourth company had described a product with our name on its own domain. An assistant that stops searching after two or three results lands on one of those and does not notice.
The fix is not on the website. It is third-party pages that say what we are, consistently, in places assistants trust. So in the three days between the two checks we did three things: published a LinkedIn company page with the same description as the site, listed our MCP server in the official registry with the same description again, and had ten guest posts go live that describe and link to us. We also kept publishing.
What the assistants said on 3 September
Right, including the operator, which it can only have found off-site. Three days from "parked domain" to a correct paragraph.
Still the wrong company, now with our build log stitched onto it. It cited our homepage among its sources and still preferred the other description. That last sentence is the interesting one: it had read enough of us to know about the log, and not enough to notice the contradiction.
Grok did not answer within the tool's time limit on the second run, which we record as no result rather than as a right or wrong answer.
What this proves, and what it does not
- It proves that a correct, crawlable website with the right files is not sufficient for AI assistants to describe you correctly, if nothing else on the web corroborates it and the name is crowded.
- It shows that off-site corroboration works fast on at least one assistant: ChatGPT corrected itself within three days of the company page and registry listing going up. We cannot say which of the three changes did it, and we did not try to isolate them.
- It does not prove anything about Claude in general. We tested one model tier deliberately, the one most users talk to. A larger model in the same family, run on the earlier date, searched more sources and said honestly that it could not find reliable information, rather than guessing.
- Two checks are two checks. We will run this monthly and add each result here, right or wrong.
Update, 4 September
Hours after this post went up, we asked Claude the same question in the claude.ai app rather than through the API, and it described rank.fast correctly: the indexing service verified through Search Console, the curated guest post pilot built from roughly 160,000 listings, the operator, and the public log. It also separated us from the three similarly named companies by name. Asked whether it would recommend us, it declined, on the grounds that everything it could find was our own description of ourselves and there were no independent reviews yet. That is a fair reading and it is the gap we have left to close. The tool keeps testing the model tier most people use, so the earlier answers stand as recorded; this note is the third data point.
Check your own
The AI visibility check is free: three checks a day, no signup, about forty seconds. It tells you which crawlers your robots.txt blocks, whether you have an llms.txt and structured data, and what each assistant says about you from memory and with search on, with the sources it used. If the answer is wrong, the fix is usually the same as ours: not the website, but what the rest of the web says about it.