We audited 701 AI tool sites, then rebuilt the check we run on yours

Over the last few days we measured every site in this catalogue four times over: the share image, the title and meta description, what the pricing page says, and how much HTML the homepage ships. 701 listings, published one article at a time.

Then we looked at our own submission check — the one a founder runs by pasting a URL into the submit form — and found it would have passed almost everything the audits caught.

So we rebuilt it. This is what changed, and why each change is there.

It asked whether the tag existed. It now asks whether it works.

The title. The old check had one question: is there a <title>? Across 673 readable sites, exactly 8 had none. Meanwhile 60 had a title that is only the brand name — Ollama, CrewAI, Browserbase — and every one of them passed.

A title of Ollama can match one search: ollama, typed by people who already know Ollama. Everyone searching run llama locally is looking for that product and being shown somebody else. That is the single most common on-page problem in this dataset, and the check was silent on it while answering a question that comes up eight times.

It now says so, and tells you what to write instead. A further 125 titles run past 60 characters, which is roughly what a result shows — those get a note too, because the part you chose to end on is the part nobody reads.

The description. Same shape. Present or absent, and 41 sites had none. But 170 run past 160 characters and get cut mid-sentence, and 37 are under 50 — some of those are excellent (Read research papers 10x faster is thirty-one characters and does more work than most of the long ones), so the note asks you to check rather than telling you off.

The share image. This is the one we had explicitly admitted to. The article measuring og:image across 701 sites ended by saying that finding the tag is the same code a submission runs, and that fetching the image and measuring it is not.

Now it fetches it. That catches:

• 14 sites whose og:image URL does not load — eight plain 404s. The tag is there, the value looks right, the file is gone, and the page still looks perfect in a browser. Nothing but requesting the file finds this.
• 6 serving SVG, three AVIF, and two returning a web page where an image should be. All validate. None produce a card.
• 18 using a square favicon as a share image, from 512×512 down to one at 64×64 — which forces the small thumbnail treatment however good the artwork is.

What it deliberately does not do

It does not fail you for any of this. Every new note is a warning. A directory that refuses a listing over title length has mistaken itself for an authority on somebody else's marketing, and you may have reasons we cannot see.

It does not use a flat 1200×630 rule for share images. We tried that in the first version of the og:image article and had to publish a correction: 76 sites called undersized that render perfectly well, twelve of them at 1200×628, which is the ratio Facebook's own guidance recommends. Below roughly 600×315 the platforms stop showing a wide card. Above it, the difference is quality, not breakage. Your 1024×537 image is fine and the check says so.

It does not need an account. Paste a URL, read the seven answers, close the tab. Submitting afterwards is a separate decision.

What it looks like

Two real results, run against the live check:

ollama.com    title  warn   The title is "Ollama" — your brand and nothing else.
granola.ai    title  pass   "Granola — The AI Notepad for back-to-back meetings"

tabnine.com   image  warn   No og:image, so links to your site share as bare text.
granola.ai    image  pass   A social share image is set: 1024×537.

Every warning carries a specific fix, not a category. "Add what the product does" with your own title in the sentence, rather than "improve your title tag".

Why we are telling you this

Because the interesting part is not that we shipped a check. It is that a check written carefully, by people paying attention, spent months answering the rare version of each question and staying quiet on the common one — and the only reason we found out was measuring a few hundred real sites and reading what came back.

If you are running a similar check on your own product, that is the transferable part. Ours passed a page whose share image was a 404, because it never asked for the file.

Run it on your site. It takes about ten seconds and we do not keep anything you do not submit.

What we are not claiming

The seven checks are worth what they measure and no more. They say nothing about whether your product is good, whether the market wants it, or whether you will rank — only whether the page is telling a machine what you meant it to.

"Brand name and nothing else" is matched against your domain, not against your product name, because the name is not known at that point. That is conservative by design: a title carrying any word the domain does not have is left alone, which means we miss some and flag none we should not.

A share image we cannot fetch is reported as not loading, which is honest about what a scraper would see and unfair to a CDN having a bad minute. It is a warning either way, and warnings cost nothing.