of answers citing their page never named them
The assistant read the source and recommended a competitor from it instead. On a citation-share dashboard this reads as a win.
The second problem
Cited, and still not recommended.
There is a kind of losing that looks exactly like winning. Your page gets cited. The assistant quotes your research, borrows your framing, uses your numbers — and then recommends somebody else.
Ahrefs published 34 self-promotional pages across five domains and read 9,886 AI answers from ChatGPT, Gemini, Perplexity and Copilot between February and May 2026. Among the answers that cited their own page, 43% never mentioned their brand at all. The model used their research and named a competitor out of it.
of an established brand’s new mentions came from third-party pages
Only 6% came from answers citing its own content. You cannot self-describe your way into a recommendation.
of eligible days a cited page actually appeared
About one day in three. Checking once and drawing a conclusion is close to guessing.
mention rate on the narrow query — against 15.8% on the broad one
Same pages, same brand, same window. Which questions you go after matters more than how many.
Ahrefs, Self-Promotional Content and AI Search, 2026. 34 pages, 5 domains, 9,886 answers, four engines, February–May 2026.
The mechanism
Why a quote and a recommendation come apart.
A retriever lifts a passage, not a page. That single fact explains most of it. When it pulls the paragraph containing your category term, it takes whatever else is in that paragraph — and if your differentiator is two paragraphs further down, it simply is not in the material being read. The model then has a category, a question, and no particular reason to attach your name to the answer. It picks whichever brand is nearest, which is frequently a competitor you named yourself in a comparison table.
The second cause is consensus. Assistants weight independent agreement far above self-description, which is exactly what the 94% figure is measuring: an established brand’s new mentions came overwhelmingly from pages it did not own. Your own site is evidence that you make a claim. Someone else’s site is evidence that the claim is true. Where those two diverge, the model resolves toward the crowd, and the crowd is currently talking about somebody else.
The tell
- Citation share rising while pipeline is flat. The classic signature. You are being read and not recommended, and the dashboard cannot see the difference.
- You appear in answers about the category but not in answers about the choice. Informational queries name you; commercial ones do not.
- Your comparison pages get cited most. They are the most quotable thing you own and the most likely to hand the recommendation away.
The fix
Five moves, in order.
None of them is “write more.” Three of them are edits to pages you already have.
- 01
Separate the two numbers
Score every prompt twice — cited, and recommended. Most tracking reports only the first, which is why this failure mode goes unnoticed for quarters. Where the two diverge you have located the problem precisely.
- 02
Read the cited passage, not the page
Open the exact paragraph the assistant quoted. Nine times in ten the category term is there and your differentiator is two paragraphs down, or a competitor's name is closer to it than yours.
- 03
Collapse the distance
Rewrite so the claim and the category live in the same sentence, ideally the same clause. Not “we do X. We are faster than most.” but “we are the fastest X for Y.” A retriever cannot split what is grammatically joined.
- 04
Audit competitor adjacency
Find every page where your name already appears and note whose name sits next to it. Comparison pages you wrote yourself are the usual culprit — you introduced the competitor to the model, and it took the recommendation from you.
- 05
Move the claim off your own site
94% of an established brand’s new mentions came from third-party pages in the Ahrefs data, against 6% from its own. If review platforms and roundups make the association for you, the model does not have to take your word for it.
Questions
Cited but not recommended, answered.
Why does ChatGPT cite my page but recommend a competitor?
Because the model lifts a passage, not a page. If your category term and your differentiator sit in different paragraphs, a retriever can take the category from you and the recommendation from whichever brand is named nearest to it — often a competitor you mentioned yourself in a comparison. Ahrefs measured this across 9,886 answers and found 43% of the ones citing their own page never mentioned their brand at all.
How do I know if this is happening to me?
Track two numbers per prompt instead of one: whether you were cited, and whether you were recommended. Most tools report only the first. Where the gap between them is wide you have a passage problem, not a visibility problem — and more content will make it worse, because it gives the retriever more of your material to quote while somebody else gets named.
Does adding more content fix it?
Almost never, and it is the most common wrong move. The failure is structural: a claim and a category sitting too far apart, or a comparison page that names three rivals in the same breath as your category term. The fix is rewriting the paragraph so the two cannot be separated, and getting third parties to make the association on pages you do not own.
How often does a cited page actually appear?
About one day in three between its first and last appearance, in the same Ahrefs experiment. Citation is far more volatile than ranking, which is why a single spot-check is close to guessing and why anyone reporting AI visibility from a monthly manual check is reporting noise.
BackgroundGEO vs SEOThe nine decisions that differ, and why ranking is weak evidence of retrievability.
ReferenceRecommendation rate, definedTwelve terms including the two numbers this page is about.
PracticalWhat this work costsPublished rates, and what to expect at every budget.
Want to know whether this is happening to you?
Send your site and I will run an automated read of it — homepage, robots.txt, llms.txt — and email you a ranked fix list. Free, and no call attached.