We checked 48 pages that promise AI citations. 26 were never cited once

What we actually ran

Five queries that people really type — from “best ai ppt maker” to “how to get cited by chatgpt” — asked six times each to the same model at the same temperature, 30 runs in total. Two of them came back unparseable and were dropped from both the numerator and the denominator, which leaves 28 usable runs. We did not replace them with a retry: a run we could not read is not a run that showed nothing.

For every run we recorded two separate things: the retrieval pool — every source the engine put in front of itself while composing — and the citation set, meaning the sources the written answer actually named. Those two sets are not the same, and the gap between them is the whole subject of this site.

The number

• 48 pages sat in the pools across all five queries.
• 22 of them were named at least once — 45.8% of the field.
• 26 were named zero times — 54.2%.

More than half of the pages competing for an AI citation never receive one, and the two reasons they do not need different fixes — which is the only thing this whole site is built to tell apart.

Two ways to lose, and they are not the same

From the outside both failures look identical: you search, the assistant answers, your name is not in it. From the inside they are opposite problems. Either the engine never had your page in front of it — in which case nothing about your writing matters, because nothing was read — or it read your page and quoted someone else, in which case your page is fine and the sentence it produced is not.

The fix for the first is mechanical: let the retrieval crawlers in, publish an llms.txt that is a file rather than a fallback page, and put structured data on the page so the engine has something to lift. The fix for the second is writing — a summary that matches the shape of the question. One of those you can do this afternoon; the other is the part nobody sells.

What the losing half looked like

We did not name them, and that is deliberate: a page that already lost should not get a backlink out of it. But the shape of the failures is worth publishing, because it is the most actionable thing in this dataset.

• 19 of the 48 pages carried no structured data at all — not badly configured, absent.
• 4 answered /llms.txt with a 200 and a page of HTML — a single-page-app fallback dressed up as a file.
• 5 blocked at least one retrieval crawler at the root, and 7 blocked at least one AI crawler of any kind. Those two counts are different and must never be quoted as one.
• 1 had no extractable body text at all — a page a crawler reaches and finds nothing on, which is the state that makes every other fix pointless.

What this does not tell you

48 pages, one model, one day, five queries. This is not a ranking study and it is not a prediction. If you change your page tomorrow, nothing here says the number will move — we measured a snapshot, and we say so on the methodology page in the same words.