Why ChatGPT reads your page and doesn't cite it
You're not invisible. You're uncited.
AI search has no leaderboard - the answer is rebuilt every query. AirOps found ChatGPT read 200,000+ pages across 7,500 commercial queries and cited only about 15%. The rest were opened, read, and set aside. That is not a visibility problem. It is a conversion problem on pages you already paid for.
Are you unretrieved, or just unpicked?
Most teams assume the models cannot find them, so they greenlight more content. That assumption is usually wrong, and the mistake is expensive: it turns a same-day formatting job into a twelve-month programme. A new page also starts at zero and has to earn its way into the retrieval pool first, so volume is the slowest possible answer to a short-term pipeline problem.
In SEO you outlast people. In AEO you outstructure them. Waiting takes years. Structure takes an afternoon. But you have to know which problem you actually have.
Three facts about how AI citations work
Roughly 60% of AI Overview citations go to pages that don't rank in the top 20 at all. Page one is not the entry fee.
Seven of the top 10 are 3+ years old. You lose on tenure, not quality - and tenure can't be outwritten. AI search builds a new line every query, and over half of brands that drop out are back within two runs.
Across 7,500 commercial queries, ChatGPT opened 200,000+ pages and used about 15%. The other 85% did the hard part - they were in the pool - and still lost the slot.
The three reasons a page isn't cited
Find which row you're in before you spend a dollar. Most teams are in row two and budget for row one.
| Diagnosis | What it means | What it looks like in practice | Cost to fix |
|---|---|---|---|
| Not in the pool | The model never retrieves you | Digital PR to the domains models already cite. Getting into "best X" listicles and roundups. Reddit and community presence. Claimed G2, TrustRadius and Capterra profiles. Original first-party research others quote. One consistent entity description across LinkedIn, YouTube and directories. | New authorityMonths. Start it now, but don't expect it to move this quarter. |
| In the pool, not picked | The model read your page and chose someone cleaner | Comparison tables, three per page. Real price numbers, not ranges or sliders. FAQ blocks and schema markup. Your densest paragraph broken into a list. One fresh first-party stat. Claims cut to 10 words. Clean H1/H2 hierarchy. Updated publish date, URL resubmitted. | StructureAn afternoon. No brief, no writer, no design queue. |
| Picked, not named | You're cited as a link, but your brand isn't spelled out in the answer | Brand name inside the extractable block itself, not just the header. Identical naming everywhere it appears. Pushing into the top three of listicles, where roughly 80% of brand mentions sit - position nine barely counts. | BothWeeks. Only 28% of answers contain both, and those brands are 40% more likely to reappear. |
Heading hierarchy, FAQs and tables get filed under "we need a content programme" and then never ship. They belong in row two. They are same-day work. Sorting your fixes into the right row is the single highest-leverage hour in this whole exercise.
How often should you update pages for AI search?
Weekly beats twice a year, permanently, because the answer is rebuilt every time someone asks. Pages untouched for three months are three times more likely to lose their citations, and 83% of citations on commercial queries go to pages updated within the last year.
The constraint is cycle time, not creativity. Chime cut production time 89% and tripled citations in the same window. Vin went from two hours a page to ten minutes and grew citations 600% with two people. Neither got more creative. They got more turns at bat. Ten pages a week teaches you what earns a citation; one page a month teaches you nothing.
The test for every change: if it needs a brief, you're doing too much. A normal refresh makes a page better for a human. This makes it extractable for a model. Different job, different edit, different week.
Caveat: this research is vendor-published and retrieval behaviour changes without notice. Treat it as a testing loop, not a playbook.