What the retrieval runs actually showed
Every post here comes from the same controlled experiment — 48 sources, five queries, six runs each — and every number in them is recomputed from the raw run records by one script. Where the data cannot support a claim, the post says so in the same paragraph that makes the claim.
We checked 48 pages that promise AI citations. 26 were never cited once
Five queries, six runs each, every source the assistant surfaced written down. Of the 48 pages in those pools, 26 were never named in a single answer. That is not a small tail — it is more than half the field.
Data · 7 min Read →Ranked first in the pool, cited zero times
10 pages entered their pool at position one. 5 of them were never named in a single answer. Meanwhile 9 of the 22 pages that were cited did not rank in the top three at all.
Data · 6 min Read →The llms.txt that returns HTML
4 of the 48 sites we measured answer /llms.txt with a 200 and a page of HTML. Any check that looks only at the status code calls all 4 of them healthy.
Method · 5 min Read →JSON-LD and citations: 36.0% vs 10.5%, and why I will not call it a cause
Pages carrying JSON-LD were cited in 36.0% of runs. Pages without were cited in 10.5%. That is the strongest page-level signal in our dataset, and it is still not a cause.
Data · 8 min Read →Which queries have no winner
Run the same query six times, then compare the set of sources each run named. Across our five queries that overlap ranged from 0.472 to 0.700 — and that single number says whether a query has an incumbent, and whether it is worth competing for at all.
Data · 6 min Read →The data behind every post
The method, the thresholds, and the caveats are published in full. If a number here disagrees with that page, that page wins.