Most people worrying about AI search are worrying about the wrong result. They picture being invisible, a total blank, the assistant answering as though the company does not exist. That is bad, and it is also the easy problem, because a zero at least tells you where you stand.
There is a worse result and almost nobody talks about it. The assistant opens your page. It reads your comparison table, your pricing, your definitions. It uses all of that to write a confident three-sentence answer. Then it names three other companies and puts your URL in the source list at the bottom.
You did the work. Somebody else got the customer. Your analytics show a citation, your traffic shows nothing much, and the shape of the loss is invisible unless you go looking for it on purpose.
- ”Cited but not recommended” is when AI uses your page as a source and names a competitor in the answer. Your URL is in the citation list. Your brand is not in the sentence.
- It usually means the page reads as reference material, not as a product. Roundups, definitions and how-to guides get quoted for their facts, and a model cannot name a product the page never states plainly.
- You cannot see it in a visibility score. A score counts mentions. Finding this needs the answer text and the source list side by side, which is what AI visibility tracking in ContextBolt SEO stores for every question, from inside the agent you already work in, on a board that keeps them, for $35 a month.
- The fix is a page edit, not a link campaign. Put yourself in the comparison you wrote, with the same specificity you gave everyone else.
- Check it question by question. The same page can be cited and named on one question and cited and skipped on the next.
What “cited but not recommended” actually is
Every AI answer has two layers, and they are produced by different parts of the system. The sources are the pages the model retrieved and read. The answer is what it wrote after reading them. Nothing forces those two layers to agree about who is worth buying from.
So there are three states your brand can be in for any given question, not two.
| State | What it looks like | What it is worth |
|---|---|---|
| Named | Your brand appears in the answer text, usually with a reason | The recommendation. This is the one that converts. |
| Cited, not named | Your URL is in the source list, your brand is not in the text | A link position and a small trickle of clicks. The intent goes to whoever was named. |
| Absent | Neither. The model did not read you and did not name you | Nothing, and at least it is obvious. |
Most tracking tools collapse this into one number. They count mentions, divide by something, and hand you a score out of 100. A score cannot tell the difference between “absent” and “cited but skipped”, and those two need completely different fixes. One needs content that exists at all. The other needs an edit to a page you already published.
Cited but not recommended is the middle state: your page is good enough to be read, and your product is not present enough in it to be named.
Why it happens to good pages
The pages this happens to are usually the best ones on the site. That is the uncomfortable part.
Reference tone
The page reads as neutral reference. If you wrote “the seven best project management tools” and reviewed everyone fairly, you built exactly the artifact a model wants to quote. Facts, comparisons, structure. But if your own product appears once at the bottom under a heading that reads like a disclosure, there is nothing there for the model to lift into a recommendation. It quotes your facts about other people.
Missing the obvious
Your own product is not on your own list. This sounds absurd and it is extremely common. Founders leave themselves off a roundup because including yourself feels like cheating. The result is a page that argues persuasively for four competitors, gets cited for it, and never once states what you do.
No reason attached
You are on the page but with no reason attached. A model names brands it can justify in a clause. “Linear, for teams that want speed” is nameable. “We also offer project management” is not. If every competitor on your page has a specific one-line case and you have a generic sentence, the specific ones get picked.
The wrong question
The page answers a different question than the one being asked. Your definition post about what a CRM is gets cited on “best CRM for startups” because it explains the category well. It was never written to answer that question, so nothing in it recommends anything.
Ahrefs’ research on where AI Overview citations come from found that most cited pages are not the top-ranking pages for the query. Retrieval and ranking are different jobs. Your page can be retrievable and unrecommendable at the same time, and the two problems have no relationship to each other.
How to spot it
You need three things together: the question, the answer, and the sources. Two out of three tells you nothing.
Ask the question the way a buyer asks it. Not your brand name, and not a keyword. “Best CRM for a five-person startup” is a question. “CRM software” is a keyword, and a keyword returns a definition rather than a recommendation.
Read the answer for names. Write down every brand in the text, in order. That list is the recommendation set for that question, and it is the only list that matters.
Then read the source list. If your domain is in it and your brand is not in the names list, you have found one. Note the exact URL, because that is the page you are going to fix.
A shortcut for the names half, not the whole test. Our free AI visibility checker reads an index of ChatGPT and Google AI answers for your category and tells you whether your brand is named and which rivals are. It does not return the source list, so the third step above is still yours to do. What it saves you is finding out whether you have anything worth investigating before you spend the afternoon.
Repeat the same question later. AI answers are not deterministic. SparkToro’s testing found that two runs of the same prompt rarely return the same brand list, so one reading is an anecdote. Three readings across a couple of weeks is a pattern. Our guide to how AI assistants choose citations covers what actually drives the retrieval half of this.
Doing that by hand for ten questions across four engines takes an afternoon, which is why it quietly stops happening after the first month. That is the honest argument for tooling, and it is the same argument as the one for rank tracking twenty years ago.
How to fix the page
The fix is almost always an edit to the page that got cited, and it is smaller than people expect.
Put yourself on the list, properly. Same format as every other entry. Same depth. If the others get a price, a “best for” line and a limitation, you get a price, a “best for” line and a limitation. A model will not promote you above your own formatting.
State the limitation out loud. This feels wrong and it works. “Not the right pick if you need offline mode” makes the rest of the entry credible, and credible entries are the ones that get lifted. Sanitized copy reads as marketing and gets skipped in favor of the entry that sounds like a review.
Say what you are in one sentence, near the top. Not your tagline. The category word plus who it is for. A model needs a sentence it can quote without editing, and if it has to construct one from three paragraphs it will usually construct one about a competitor instead.
Answer the question the page is being cited for. If your definition post keeps getting cited on a “best X” question, that is the search telling you a page is missing. Write the page that answers it. This is the single highest-value signal in the whole exercise, because the model has already told you which question it thinks your site is qualified to answer.
Then check it again, on the same question. AI answers rebuild from an index that refreshes continuously, so this moves in days and weeks rather than the months a ranking takes. That is the one genuinely good thing about this channel.
What not to do
Do not stuff your brand name into the page. Repetition is not a citation signal and it makes the page worse to read, which makes it worse to quote.
Do not remove the competitors. The reason your page got read is that it was useful about the whole category. Strip that out and you lose the citation as well as the recommendation, and you end up with a brochure nobody retrieves.
Do not chase a single bad answer. Run the question a few times before you rewrite anything. Ask two different engines. The variance between runs is larger than most of the changes you are considering making, and rebuilding a page around one unlucky reading is how you waste a week.
Do not treat the score as the work. A visibility score going from 12 to 19 tells you almost nothing about which page to open, and the score itself is not comparable between tools anyway. The buyer complaint about this whole category, quoted repeatedly through 2026, is that the tools benchmark and nothing changes. A number with no page attached is exactly that.
How common is this?
Common enough that it is the normal result rather than the edge case.
When we first ran the full report on our own site, both of the indexed questions naming us came back this way. Cited as a source, and recommending somebody else in the same sentence. We had written the comparison content that taught the model about the category, and the category answer named a competitor.
That is not a complaint about the models. It is a description of how retrieval-augmented answers work. The system rewards pages that are useful about a subject, then recommends whichever product on those pages is stated most clearly. Those are two different jobs and most sites only did the first one.
If you want the measurement running without the afternoon of manual checking, AI visibility tracking inside ContextBolt SEO stores the question, the full answer and every source together, so the “cited but not named” cases are visible instead of averaged away. It runs from Claude, Cursor or Codex for $35 a month, alongside keyword, SERP, backlink and Search Console data, which means the page you need to rewrite and the thing that rewrites it are in the same conversation. Google’s own generative AI performance reports give you impressions but not this, so the source list is still something you have to go and read.
Start with your best comparison page. It is the one most likely to be cited, and most likely to have left you off its own list.