Last updated: July 25, 2026 · LeadsNow AI, Melbourne
At a glance: ChatGPT cites only 15% of the pages it retrieves — the other 85% get read and dropped. Retrieval gets you into the room; extractability wins the citation. To make the 15%, lead every section with a direct one-to-three-sentence answer, cover the fan-out sub-questions ChatGPT actually asks, and keep the page visibly fresh.
Here is the part nobody tells you when ChatGPT ignores your site: it has probably already read your page. In March 2026, AirOps analysed 548,534 pages ChatGPT retrieved across 15,000 prompts and found only 15% of them were cited in the final answer.
Sit with that for a second. Eighty-five per cent of the pages ChatGPT pulls into an answer get read, weighed and binned — without the searcher ever knowing they existed. Being retrieved is not the prize. It is the audition.
This guide is about the last mile: what separates the 15% that get credited from the 85% that do the work and get nothing. If you want the wider per-engine picture, we covered why ChatGPT and Perplexity cite completely different sources separately — this piece is purely about closing the retrieval-to-citation gap in ChatGPT.
Retrieval is not citation: the 85% problem
When someone asks ChatGPT a question, three things happen in sequence:
- Fan-out. ChatGPT breaks the prompt into internal sub-queries. In the AirOps dataset, 89.6% of searches triggered two or more fan-out queries, expanding 15,000 prompts into 43,233 total searches.
- Retrieval. It pulls candidate pages for each sub-query — over half a million pages in the study.
- Citation. It writes the answer and credits a shortlist. Only 15% of retrieved pages make that cut.
Most “AI SEO” advice stops at step two: get indexed, get ranked, get retrieved. But steps one and three are where citations are actually won. Ranking helps — pages at #1 in Google were cited 43.2% of the time, 3.5 times the rate of pages outside the top 20 — but flip that number around: even the #1 ranking page misses the citation more often than it earns it. Rank alone does not close the deal.
Retrieved-but-ignored vs cited: what the 15% do differently
| Trait | Retrieved but never cited (the 85%) | Cited (the 15%) |
|---|---|---|
| Answer placement | Answer buried mid-page after preamble and throat-clearing | Direct answer in the first one to three sentences under the matching heading |
| Query match | Targets the broad head term only | Resolves the specific fan-out sub-question, not just the original prompt |
| Extractability | Claims spread across long paragraphs; nothing quotable in isolation | Self-contained claims — a sentence lifted out still makes sense, with the number and the noun in it |
| Evidence | Opinion and adjectives | Named data, dates, sources — something the model can attribute |
| Freshness | No visible update date; content aging quietly | Updated within the last 12 months, with the date visible — 83% of commercial-query citations come from content updated in the past year |
| Job on the page | Tries to cover everything about the topic | Does one job per section, cleanly — easy for the model to slot into one answer sentence |
The fan-out loophole: a third of citations come from queries nobody tracks
The most useful finding in the AirOps study is not the 15%. It is this: among cited pages that appeared in any top-20 SERP, 32.9% were discovered only through fan-out queries — they never appeared in the results for the original prompt at all.
And here is the kicker: 95% of ChatGPT’s fan-out queries had zero monthly search volume by traditional metrics. Your keyword tool calls them worthless. ChatGPT is running them thousands of times a day and handing out citations to whoever answers them.
That is a genuine loophole for smaller sites. You may never outrank the incumbents for “lead generation agency Australia”. But nobody is competing for the fan-out sub-questions behind it — the response-time comparisons, the pricing-model breakdowns, the “does X work for Y industry” questions — because they show zero volume and every volume-led content plan skips them.
How to find your fan-out queries
- Ask ChatGPT your money questions and watch what it searches. The searches it runs mid-answer are shown in the interface. Those are your real target queries.
- Decompose every head term. For each commercial query you care about, write the five to ten sub-questions a careful buyer would need answered: cost, comparison, timeframe, proof, fit-for-my-industry.
- Give each sub-question its own H2 or H3 — phrased as the question — with the answer immediately underneath. One section, one job.
Extractability: writing answer-shaped blocks
ChatGPT does not cite pages. It cites passages — the specific block that resolves the sub-question it is holding. So the unit of optimisation is the block, not the article. The pattern that works:
- Heading = the question, in the words a buyer would use. Not clever, not branded. “How fast should you call a new lead?” beats “Speed wins”.
- First sentence = the answer. Complete, self-contained, with the number in it. If that sentence were quoted alone in a ChatGPT answer, it should still be true, attributable and useful.
- Next two to four sentences = evidence and nuance. The data point, the source, the exception.
- Then stop. Start the next block. A 3,000-word page can absolutely earn citations — as a stack of twenty answer-shaped blocks, not as one essay.
Add the supporting furniture: a summary capsule up top (like the one on this page), comparison tables for anything with two options, FAQ markup that mirrors the visible questions, and an honest “last updated” date. None of it is exotic. It is simply making every claim liftable.
This is exactly how we build our own pages and our clients’ — and then we verify it worked by tracking share of answer rather than rankings. If you would rather see the before-and-after on real pipeline than read about it, Book a call.
Freshness: the quiet filter on commercial queries
Extractability decides who wins the passage. Freshness decides who stays in the pool. AirOps’ stale-content research found 83% of citations for commercial queries come from content updated within the past year, 60% from content updated within six months — and pages left unrefreshed for over 12 months are more than twice as unlikely to be cited.
The practical move is a refresh cadence, not a rewrite habit: put your money pages on a quarterly rotation, update the numbers and dates that have actually changed, and surface the new date on the page and in your schema. A stale top ranker loses to a fresher, lower-ranked page in this game more often than SEO instinct suggests.
What this looks like when it feeds a pipeline
We care about the 15% because our own demand runs through it. LeadsNow’s pipeline is built on being the page AI engines quote when Australian businesses ask about lead generation and appointment setting — the same discipline described above, applied to our own funnels and our clients’. That sits on top of 50,769+ AI-booked sales appointments since 2017 and 1M+ leads generated, with 25 filmed client case studies behind it. Citations are the input; closed deals are how we score it.
It also compounds with authority work: extractable pages earn the citation, and earned media builds the trust that gets you retrieved in the first place. The same answer-shaped blocks also travel to other engines — with different weighting, as we showed in how B2B brands get cited in Perplexity. Do both and the 15% stops being a lottery.
Want to know which of your pages are being retrieved and dropped? We will map your money questions, show you where you are losing the last mile, and what fixing it is worth in booked appointments. Book a call.
Frequently asked questions
Why does ChatGPT not cite my site even though it can access it?
Almost certainly because you are being retrieved but out-answered. AirOps’ analysis of 548,534 retrieved pages, covered by Search Engine Land, found only 15% of pages ChatGPT retrieves ever appear in the final answer. The citation goes to the passage that most directly resolves the specific sub-question — so the fix is answer-first formatting on the exact fan-out questions, not more indexing or more backlinks.
Does ranking #1 in Google guarantee a ChatGPT citation?
No. In the same AirOps dataset, pages ranking #1 in Google were cited 43.2% of the time — 3.5 times more often than pages outside the top 20, but still a miss on most answers. Ranking gets you retrieved reliably; the citation is decided by how extractable and current your answer is once ChatGPT is reading you against the rest of the shortlist.
Should I target keywords with zero search volume?
For AI citations, yes — deliberately. 95% of ChatGPT’s fan-out queries had zero monthly search volume, and roughly a third of cited pages that appeared in any top-20 SERP were found only through those fan-out queries. Zero-volume sub-questions are where smaller sites beat incumbents, because volume-led content plans never touch them.
How often should I update pages I want ChatGPT to cite?
At least annually, ideally quarterly for commercial pages. AirOps found 83% of citations on commercial queries come from content updated within the past year, and pages unrefreshed for over 12 months are more than twice as unlikely to be cited. Update real substance — numbers, dates, examples — and show the date on the page.
Is being in the 15% worth the effort for a lead-generation business?
Yes, because AI-referred buyers arrive pre-sold: they have already read a synthesised comparison and clicked through to the source the engine trusted. The way to know for your own business is to measure share of answer on your revenue-critical questions, then track what those visitors do — we score it in booked appointments and closed deals, not impressions.