Our methods, our samples — including the results that flatter nobody.

Question framing decides whether retrieval happens at all

In the pre-registered ladder, none of the conceptual “what is X” framings triggered source retrieval on the engine we measured, while every dated, comparison-shaped framing did. Limit: five topics, one engine primarily; engines differ — Perplexity, for example, cites far more freely. The direction was unambiguous; the exact rates are not portable.

Domain authority did not decide citations

Across 21 matched pairs, the higher-authority domain won 11 and lost 10 — statistically nothing. Our weekly sweeps keep finding small, recently created sites holding citation slots beside household names. Limit: 21 pairs cannot rule out a small effect; they rule out a decisive one in this sample.

The unit of citation is the passage, not the page

Some assistant citations link to a single highlighted sentence inside a page rather than the page itself — one article in our harvest was cited four times for four different passages. This is why our editor measures sentences an AI could lift, not pages.

Repeating what already ranks is not what gets cited

What we measured. Across the 33-keyword harvest we recorded 276 ranking pages and 73 pages ChatGPT actually cited for those same queries. Only 9 pages appeared in both sets — 3.3% — and just 12.3% of citations came from Google’s top ten. Seven citations in eight came from outside the first page of results. The two funnels barely overlap, so a page assembled from the consensus of everything already ranking is competing in the wrong set.

What the engine says about itself. Asked to describe its own source selection, ChatGPT told us it often ignores pages that add no new evidence — five articles repeating one press release are not five independent sources. That is a self-description, not a measurement: models confabulate plausible mechanisms, and we would not publish it on its own. We publish it because it matches what we measured first, and because it named the defect in our own product — our early scorer rewarded matching the top-ten consensus, which is the set our harvest shows is nearly disjoint from what gets cited.

Limits. 33 keywords, three related industries, mostly ChatGPT. The overlap counts are pages, not domains, and one harvest is one moment — these are the odds we observed, not a rule about the web.

Harvest and engine self-report measured 29 July 2026 · published here 12 August 2026.

The null we published against ourselves

We believed answering the questions searchers ask (Google's People-Also-Ask) would earn citations. It did not: across the 33-keyword harvest, cited pages answered 58.8% of searcher questions and ignored pages answered 58.2% — a difference of nothing. We kept the measure in our editor because it remains our strongest signal of article QUALITY against the blind judges’ ratings (15 articles) — but we stopped claiming it buys citations, and our product copy says so.

Our own content score pointed the wrong way

The finding. In a pre-registered head-to-head trial, 15 articles were each read cold by an independent reviewer who saw no scores at all. Our fitted content score ran negativeagainst those judgements: rho −0.61 across the 9 articles our score covered. The competitor score we benchmarked did the same thing on its 6 (rho −0.63), and combined across all 15 the figure is again −0.61. Inside a keyword pair, the scores named the judge’s winner once in six tries: one draft scored 90 and drew a judge’s 81, while a 78 drew an 86. This is a category-wide flaw in scores fitted to what ranks — ours included, and ours first.

What the two sample sizes mean. Both numbers in our records are real and count different things: 15 is the number of articles blind-judged; 9 is how many of them carried our score (the other 6 carried the competitor’s, because each article was written under one stack or the other). So rho −0.61 (n=9) is our meter against the judges, and rho −0.61 (n=15) is the whole trial, each article measured against whichever score it carried.

Scope — read this part. Every draft in the trial scored between 71 and 90. What failed was not the meter’s ability to tell bad writing from good; it was the meter’s ability to rank good writing against itself. Small n, one narrow band. A confound we logged rather than buried: our editing loop revised each draft until the score cleared 75, so the weaker-scoring drafts received more editorial work than the strong ones. The cold reads were done by independent reviewer agents, one per article, on a seven-dimension scorecard; a second format — a three-lens side-by-side panel over 9 pairs — agreed with the first on 7 of those 9. A human blind read of the same articles is prepared and not yet done.

Why it is on this page. It is the reason we stopped selling a score as a quality grade and rebuilt ours as a coverage checklist — what you are missing, not how good you are. Any tool that fits a number to the consensus of the ranking pages should be asked for this measurement. We ran it on ourselves and published the answer.

Measured 28 July 2026 · published here 12 August 2026.

Where real winners score

The 43 pages ChatGPT actually cited score a median of 65 on our citability checklist — and only 5% reach 85. Real published professional articles score a median of ~76 on our SEO-terms meter (44 measured). This is why our score bands call 55–84 “strong” and treat 85+ as informational, never a target: chasing points past the winners’ zone produces the stuffing assistants discard.