Last week I measured how often AI search engines mention or cite my company across 28 questions a buyer might ask. The public score was zero out of 28. Three studies published in late October explain why, and the explanation is not "write better pages." Being cited is two races multiplied together: first your page has to make the shortlist, then it has to be chosen from inside it. The second race is decided by how many other sites vouch for the page and whether it agrees with the consensus — while formatting, schema, and the much-hyped llms.txt file perform at roughly the level of a coin flip.
Start with the shortlist, because it is the only round most GEO advice talks about. A page gets retrieved if it is indexed, matches the topic, is fresh, and — on Google's surfaces, as reported in summaries of Ahrefs' October 29 stability study — already ranks for the question; pages that rank were about seven times more likely to still be cited a week later. That round is real, and structure helps you qualify for it. A study of 96,418 URLs by SISTRIX, also read via summary, found in-domain advantages for things like article schema and clear sectioning. Fine. That is the entry ticket. Almost everyone in this industry is polishing the entry ticket and calling it the race.
Inside the shortlist, nobody is grading your formatting
The second round works differently. A shortlist study by Cognova — a vendor study, small at 440 answers, so treat it as directional — compared pages that were read but skipped against pages actually cited within the same answer. The page with more referring domains won 62% of the time on ChatGPT. The page that agreed more with the other sources in the set won 54% of the time. Formatting, tone, and persuasive polish sat at about 50% — no effect at all. The median cited page had four referring domains and had been updated within 14 days; the median skipped page had zero referring domains and had not been touched in 30. Pages written in an all-about-us voice won just 46% of the time: inside the shortlist, self-narration reads as advertising, not as a source.
That is where my zero actually came from. It was not mainly a page-quality zero. Almost no third-party page anywhere vouches for what my company does, and no other source repeats our claims, so an engine cross-checking the shortlist has nothing to confirm us with. Choosing someone else is the rational move. You cannot format your way out of having no witnesses.
The llms.txt story is the cleanest proof of the misallocation. Google's own guide to generative search features — as reported consistently across multiple summaries of the page — says Search simply ignores the file. Ahrefs measured 137,000 domains and found 97% of valid llms.txt files received zero requests in an entire month. SISTRIX found no citation advantage either: 11.7% for pages on llms.txt domains versus a 14.0% baseline. Keep the file if you like; it costs nothing and works as a readme for visiting agents. But the industry spent a year scoring it as a citation asset. It never was one.
And yesterday's citation is already half gone
The last illusion is measurement. In the Ahrefs tracking — 7,831 questions re-asked daily for 20 days — between 50% and 83% of cited sites, depending on the platform, were gone the next day. Only 4% to 7% stayed cited for all 20 days, though about three-quarters of the disappeared ones came back. A citation is not a ranking you win and keep. It is a seat that gets re-drawn daily. Which means a one-day visibility report — the kind most teams screenshot for leadership — is statistically noise, and mentions, which churn far less than citation links, deserve to be counted separately.
So I changed three things this week. Measurement moved to a seven-day rolling window, split by question type — factual, evaluative, comparison, category — because those are different races with different natural winners. Production reordered: third-party, citable fact pieces now come before any more on-page decoration, and every answer page leads with the consensus layer — answering the category question the way the currently cited sources answer it — with our unique facts as a separate, clearly marked increment, not as the whole page in our own voice. And llms.txt was struck from our GEO scorecard entirely.
One calibration to end on. A 100-site, 12-month measurement put ChatGPT sessions at roughly 0.62% of Google organic sessions. This race is worth running as insurance — so that when someone, or some agent, checks who you are, the record is right and corroborated. It is not a growth channel, and it should never be budgeted as one. Win the second round first: get vouched for, agree where the consensus is right, add what only you know — and stop redecorating the entry ticket.
— Jiahui Miao, founder of vertciti, building a personal AI operating system for one.
