The scores
Every site was scored the same way: fetch the homepage, robots.txt, llms.txt and sitemap.xml — four requests, no rendering, no crawling — then grade the citation-readiness signals that can be verified from those four responses. You can reproduce any row below, against us or against anyone, in about five seconds.
- trakkr.ai — 100. Nothing failing.
- promvia.app — 95, and 82 when the run started. Missing: author and date signals, a visible recent year.
- otterly.ai — 92. Missing: scannable structure, author and date signals, external citations.
- rankscale.ai — 88. Missing: a canonical link, provenance signals, author and date signals.
- peec.ai — 73. Missing: organization schema, FAQ schema, entity links, content depth, author and date signals, llms.txt.
- tryprofound.com — 73, and no score at all until we fixed our own fetcher the same day. Missing: FAQ schema, entity links, author and date signals, a visible recent year, llms.txt.
A score here is not a verdict on a product. It measures how ready one page is to be quoted — not whether the company behind it is any good, and not whether assistants actually recommend them. That second question is a different measurement entirely, and it is the one our citation tracking exists for.
We published our losing score on purpose
Promvia opened at 82 — a B, and fourth of the five sites that answered. Three of the five signals we were failing are fixes this tool hands customers: our homepage carried no list markup at all, our organization schema had no entity links even though the code that builds it has always accepted them, and there was no FAQ on the page.
We fixed all three, and only those three. The numbered steps on our homepage became an ordered list, which is what they always were; the schema now names the two profiles that identify us; and the homepage carries a real FAQ — five questions people actually ask before paying, opening with whether we can guarantee an AI recommendation, answered no. That took 82 to 95.
We left two points on the table deliberately. The checker rewards a visible recent year, and pasting one onto a page to satisfy a freshness test is decoration, not freshness. If you take one thing from this table, let it be that a score is a proxy — and the moment you optimise the proxy instead of the page, you have started producing exactly the thin, gamed content assistants are learning to skip.
Nobody's homepage has author or date signals — and that is fine
One signal failed on four of the five sites we could score: author and date signals, the E-E-A-T markers that tell a reader and a crawler who wrote something and when. It is tempting to read that as five companies making the same mistake. It is not.
A marketing homepage has no author and no publication date, because it is not an article. Bolting a byline onto one would be the same decoration as the year. The signal is real — it just belongs on the pages that carry claims worth attributing.
That is measurable too. Our own homepage carries no author or datePublished, and every one of our guides carries both — the page you are reading names its date at the top and its author in its structured data. If your homepage fails this check, the useful response is to look at your articles, not your landing page.
The measurement found a bug in the measuring instrument
Profound came back unreachable, and their site is plainly up. It answers a redirect from the bare domain to the www one, and our fetcher refused every hop that changed hostname — so a very large share of the web was unauditable by us, and we had not noticed. With the hop allowed, it scores 73.
It was not only the free tool. The paid audit made the identical call, which means a customer whose domain redirects to www could have had their audit come back as a dead site. We fixed the hop the same day, narrowly: only the www label may be added or removed, so a redirect to an unrelated host is still refused and a score can never be attributed to a site nobody asked about.
The half worth copying if you build these: following the redirect alone would have been worse than the bug. The well-known files were still being fetched from the original hostname, so llms.txt and sitemap.xml would have come back missing from a host that no longer serves the site — a confidently wrong score in place of an honest failure. Ask the host that answered, not the one that was typed.
What a checker like this can and cannot tell you
Readiness is not recommendation. Everything in the table is a property of a page: is it indexable, is there schema, can the named crawlers reach it, is there anything on it worth quoting. None of it measures whether ChatGPT names you when someone asks for a recommendation, because that depends on the wording, the market, the day, and a great deal that is not on your site at all.
So treat a score like this the way you would treat a page-speed score: a fast page is not a good page, but a slow one has a specific fixable problem. Read how AI chooses sources for the other half, and what actually gets you cited for the work that follows.
Frequently asked questions
Is this table fair to the other tools?
It measures one page each, with one checker, on one day, and we included ourselves at our worst rather than after the fixes. Every row is reproducible: run any of these domains — ours included — through the free checker and you will get the same signals. What it is not is a product review; a company can score 73 and still be the right tool for you.
Why did Profound nearly get no score?
For most of the day it had none: their site redirects the bare domain to the www one, our fetcher refused the hop, and we recorded no score rather than a zero. That was our bug, not theirs. We fixed it the same day and the site scores 73. The history stays in the article because a measurement that had to be corrected is worth more to a reader than a table that never admits one.
Does a score of 100 mean AI will recommend that company?
No. It means every readiness signal this checker can verify from four HTTP requests is in place. Whether an assistant names them in an answer is a separate question, measured separately, and no tool — ours included — can guarantee the outcome.
Should I be optimising for this score?
Optimise for the underlying thing. Lists, schema and crawler access help because they make a page easier to quote correctly; a visible year helps only if the page really is current. The moment a fix exists to move a number rather than to improve the page, it has stopped working on the thing the number was standing in for.