First-party data · 12 quotable numbers
50,072measured AI citations, counted one day at a time on my own sites
Most AI SEO statistics are survey answers about what marketers believe. These are measurements: every number below comes from Bing Webmaster Tools AI Performance reports on websites I own, a 751-page controlled test on one identical template, or live citation probes against ChatGPT, Claude, Gemini, and Perplexity. Quote any of them; the attribution line is one click away, and the method sits next to every figure. The practice goes by several names: AI SEO, answer engine optimization (AEO), generative engine optimization (GEO). The numbers don't care what you call it.
Daily AI Performance exports from Bing Webmaster Tools, summed across seven verified properties, January 17 through August 5, 2026. The counter is a floor: Bing's instrument only sees Microsoft's AI surfaces (Copilot, Bing AI), so citations on other engines aren't counted here.
Bing-reported AI citations compared against AI-source referral sessions in analytics over the same window. If you measure AI visibility by watching referral traffic, you'll conclude it isn't happening while an engine cites you a hundred times a day.
751 live pages on one identical template, so topic and depth vary while design, schema, and structure hold constant. Demand odds ratio 4.42 (p = 0.003); word count odds ratio 1.09, statistically indistinguishable from no effect. Observational regression on one site, not a randomized trial.
Same 751-page single-template test. The deepest content tertile was the graveyard: depth without demand earned nothing. "Write longer, more detailed content" is the most expensive advice in AI SEO.
Page-level rollups from the same Bing instrument. The portfolio's biggest winner earned its first 6,000 citations as a client-rendered JavaScript page with no structured data. Schema is fine hygiene for classic SEO; in this dataset it did not select the winners.
875-page audit across my own sites: 99 percent of never-cited pages carry the standard formatting checklist (FAQs, question headings, schema) versus 66 percent of cited pages. The checklist doesn't hurt; it just isn't what gets a page picked.
Same template, same effort: 47 percent of pages on actively-searched topics earned citations versus 6 percent on topics without search demand. If the AI can answer from memory, it never searches, and nobody gets cited. The selection happens before your page is ever fetched.
When the game my stats site covers shipped a new season, every data page regenerated the day the update dropped. That week (ending July 12, 2026) added 8,859 citations and put the new season page from nonexistent to the site's second most-cited page. Freshness is the one checklist item the industry gets right.
Monthly sums from the daily series: April 3,202 · May 2,809 · June 8,372 · July 31,324. The curve compounds when demand windows land on pages that already rank. It also cools: the weeks after the July peak ran about 21 percent down week-over-week. Harvest, not annuity.
Across four sites: a brand-new domain with IndexNow publish pings earned its first citation in four days; new pages on an established site took six to eight; a site with no pings and weak search presence waited 64 days. Crawl setup is most of the difference, and it's free.
First citation in the series: February 22, 2026. Crossed 50,000 on August 5, 2026. The ramp wasn't gradual: it sat near zero for six weeks, climbed through spring, then compounded when demand windows hit ranked pages.
Bing's page-level reporting concentrates the portfolio's wins on 58 pages, and the top three (a game meta-rankings page, a season page, and a scholarship answer page) carry about 13,400 citations between them. Winners keep winning: densify what already earns before publishing something new.
Every number on this page is quotable with attribution; a link to this page or the linked research cut is the whole ask. Each stat above has a one-click quote with the attribution built in, or use the general citation:
Journalists and researchers: the underlying daily exports, method notes, and the evidence-tagged rulebook are public in the get-cited-by-ai repository. If you need a cut of the data that isn't published, ask; the answer is usually yes.
The counter is Bing Webmaster Tools' AI Performance report (free, first-party, per-page), summed daily across seven verified sites: a gaming data site, a college financial-aid site, a family storytelling site, an AI answer-checker, an AI prompt library, this site, and a product site. The 751-page test holds template constant so topic and depth are the only variables; it is observational, not randomized. Live probes against ChatGPT, Claude, Gemini, and Perplexity are run repeatedly, never once, because single-run probes are weather.
The limits: the instrument is Bing-heavy, so ChatGPT/Perplexity/Claude citations are measured by probe rather than counter. It's one operator's portfolio in five niches, US/English: a large, clean sample of my sites, not a census of the internet. Where a finding is untested, the research says so instead of rounding up to certainty. That discipline is the product: predictions get registered with refutation conditions before outcomes are known, in a public, git-timestamped ledger.
The seven-step playbook behind these numbers: topic selection, search presence, liftable answers, refresh cadence, measurement.
Every rule evidence-tagged from the measured dataset, with the prediction ledger and audit modules. MIT-licensed on GitHub.