Being Cited Is Not Being Recommended: The Gap AI Visibility Reports Miss

An AI answer can link to your page and still tell the reader to buy from someone else. That happens more often than most reports suggest. Your page supplies a useful fact, the AI shows your link as a source, and then the answer names a competitor as the best choice.

The reverse happens too. Some brands get named and recommended without a single link to their own site. If you only count citations, you will misread both situations. This piece explains why the gap exists, shows what it looks like in one example answer, and gives you a way to measure and test it properly.

Mentioned, cited and recommended are three different things

People often use these words as if they mean the same thing. They don't. Here is what each one means in this article.

• Mentioned. Your brand name appears in the text of the AI answer.
• Cited. A link to your website appears as a source for the answer. A citation is the AI's way of saying "some of this came from here."
• Recommended. The answer tells the reader to choose you, or ranks you as a top option for what they asked.

A fourth idea sits on top of these: position. If an answer lists four brands, being named first is usually worth more than being named fourth. So "recommended" is not a yes or no question. It also matters where you land in the list.

These four things can move separately. You can be cited but not mentioned. You can be mentioned but not cited. You can be mentioned and cited but only as a warning ("cheaper, but missing key features"). Each combination tells a different story about your brand.

How big is the gap? What one large study found

The best public data on this comes from a Semrush study published in June 2026.¹ The team logged 3,981 times a website showed up in AI answers. These came from 115 prompts run in 14 countries across four AI search tools: ChatGPT, Google AI Overviews, Gemini and Google AI Mode.

Each appearance was sorted into a bucket. "Cited" meant the website showed up as a source link. "Mentioned" meant the brand name appeared in the answer text. Here is what they found.

The study calls that first bucket a "ghost citation." In other words, in almost 62% of cases, the website was cited as a source but the reader never saw the brand's name in the text. Only about 13% of appearances got both.¹

The tools behaved very differently from each other:

• ChatGPT cited brands in 87% of its appearances but named them in only 20.7%.¹
• Gemini did almost the opposite. It named brands in 83.7% of appearances but showed a citation link only 21.4% of the time.¹

The type of question mattered as well. Informational questions had an 89.3% citation rate but only an 18% mention rate. Comparison questions had a 43.3% mention rate, which is 2.4 times higher than informational ones.¹ This makes sense. When someone asks "what is X," the AI needs facts, so it borrows them and links the source. When someone asks "which X should I pick," the AI needs to name options.

A note on the data: the study does not say when the answers were collected, and it does not give sample sizes for each tool or country. So treat the per-tool numbers as a strong signal, not an exact measurement of how each tool behaves today. The main finding still holds: citations and mentions are not the same metric, and one cannot stand in for the other.

Why the gap exists: AI answers read more than they show

To see why a page can be cited without being recommended, it helps to know how AI search tools build an answer.

Step one: the AI runs its own searches in the background

When you type a question into Google's AI Mode, it does not run just one search. Google says AI Mode uses a "query fan-out" technique. It issues multiple related searches at the same time across subtopics, then brings the results together into one answer.² Google later said AI Mode breaks your question into subtopics and runs many searches on your behalf. Its Deep Search feature "can issue hundreds of searches" for a single question.³

ChatGPT works in a similar way. OpenAI's help page says ChatGPT search usually rewrites your question into one or more targeted searches and sends those to its search partners. It may then send follow-up searches after looking at the first results.⁴

This process is called grounding. The AI "grounds" its answer in pages it just found, instead of relying only on what it learned during training.

Step two: the visible citations are only part of what was read

Because the AI ran many searches, it saw many pages. The links you see under the answer are only some of them. OpenAI's help page says the Sources panel shows "cited sources and other relevant links," which tells you the set of pages involved is bigger than the set shown as citations.⁴

Fan-out also explains why AI citations often don't match normal search rankings. An Ahrefs study of 15,000 long-tail prompts found that only about 12% of URLs cited by AI assistants ranked in Google's top 10 for the original prompt.⁵ The figure was 28.6% for Perplexity and 8.0% for ChatGPT's in-text citations. Ahrefs suggested query fan-out as one reason: the assistant searched for other versions of the question, not just the one the user typed.⁵

Step three: a citation link is a loose signal

Here is the part that surprises people. A citation does not prove that your page shaped the part of the answer that matters to you.

Researchers at Stanford University checked answers from four AI search engines in 2023 (Bing Chat, NeevaAI, Perplexity and YouChat), using 1,450 questions for each. They found that on average only 51.5% of generated sentences were fully supported by their citations. And only 74.5% of citations actually supported the sentence they were attached to.⁶ That study is a few years old, and the tools have changed since. But it shows something that still matters: the link sitting next to a sentence is not always the true source of that sentence.

Independent journalism research found the same kind of problem. The Tow Center for Digital Journalism at Columbia tested eight AI search tools with 1,600 queries in early 2025. The tools were given short passages from news articles and asked to name the original article, publisher, date and link. Together, they gave incorrect answers to more than 60% of the queries.⁷ Some tools linked to copies of articles on other sites instead of the original publisher.⁷ So even when you are the true source, the credit can land somewhere else.

Step four: the recommendation often comes from somewhere else

Put these steps together. The AI reads many pages. It takes a fact from one (and maybe links it). Then it decides which brand to recommend based on the overall picture it has formed. That picture can come from reviews, comparison articles, forums, news stories, and what the model learned during training.

Ahrefs studied 75,000 brands to see what was linked to brands being mentioned in Google's AI Overviews. The strongest factor was branded web mentions, meaning how often the brand is talked about across the web, with a correlation of 0.664. Number of backlinks had a much weaker correlation of 0.218.⁸ This is a correlation, so it does not prove cause and effect. But it fits the pattern: being talked about widely seems to matter more for being named than having one strong page.

Also, not every answer runs a search at all. OpenAI says ChatGPT may search the web automatically when a question would benefit from current information.⁴ When it doesn't search, there are no citations, but the answer can still name and recommend brands it learned about in training. That is one way a brand gets recommended without being cited.

A citation tells you the AI used your page. It does not tell you the AI chose your brand.

An illustrative composite: one prompt, four brands

Please read this first: the example below is an illustrative composite. The brand names are invented. It is not a real client, a real answer we captured, or a real study. We built it from patterns that show up across AI answers in general, to make the gap easy to see.

Say a buyer types this prompt into an AI assistant:

"What's the best project management tool for a 20-person marketing agency?"

The assistant answers with something like this (shortened):

For a 20-person agency, Kestrelly is the strongest pick. It puts client approvals and time tracking in one place. Tools in this category usually cost around $10 to $12 per user per month for teams this size [1]. Tamberlo is a good second choice if you need detailed resource planning [2]. Oxlade is cheaper, but it lacks a client portal, which most agencies will want.

Sources: [1] fernwick.com/blog/project-management-pricing-2026 [2] tamberlo.com/features/resource-planning

Now look at what each brand actually got from this answer.

What happened here

• Fernwick is a ghost citation. Its pricing guide supplied the one hard number in the answer. But the reader never sees the name Fernwick, and the answer sends them to Kestrelly.
• Kestrelly won the prompt with zero citations. The AI's picture of Kestrelly probably came from reviews, comparison articles and forum threads it read during fan-out, or from its training data.
• Tamberlo got the best all-round result: named, linked and recommended, though in second place.
• Oxlade was named, but in a way that may push buyers away. A simple "mention count" would score this as a win.

How a citations-only report would read this answer

Now suppose each brand only tracks citations. Here is what their reports would say.

Every single row is wrong in some way. Fernwick's team might celebrate. Kestrelly's team might panic and start rewriting pages that are already working. And nobody at Oxlade would learn that the AI is repeating a weakness about their product.

This is one prompt. Across a few hundred prompts, the same errors stack up. That's why a dashboard that adds up citations can point a team in the wrong direction for months.

Why winning more citations is not automatically good for business

Once teams learn that AI tools borrow facts, the obvious move is to make facts easier to borrow. Put the key number at the top of the page. Add a clear definition. Add statistics and quotes. These changes often do win citations.

There is research behind this. The 2023 research paper that introduced the term "Generative Engine Optimization" (GEO) tested ways of rewriting content to get more visibility in AI answers. The authors reported that some methods could boost visibility by up to 40%.⁹ Adding statistics and adding quotations showed the strongest gains. Keyword stuffing did not perform well.⁹

But look closely at what that paper measured. "Visibility" was measured mainly as how much of the AI answer was attributed to your source, weighted by where it appeared, plus a subjective impression score. It did not measure whether your brand was recommended, whether anyone clicked, or whether anyone bought. It was tested on a research set of questions and on Perplexity, where the authors saw gains of up to 37%, not on live business pages over months.⁹ That is useful science. It is not proof that citation tactics grow revenue.

Here are three reasons a citation win can be a business loss or a wash:

1. You may be giving away the answer. If your page's best fact appears inside the AI answer, the reader may not need to visit you. Pew Research tracked the browsing of 900 U.S. adults in March 2025. When a Google search showed an AI summary, users clicked a traditional search result on 8% of visits. Without a summary, that figure was 15%. Users clicked a link inside the AI summary itself on just 1% of visits.¹⁰
2. The fact may help a competitor's recommendation. As the composite showed, your fact can become the supporting evidence for someone else's top spot.
3. Changes for AI can affect normal search. Google says there are no extra requirements or special optimizations needed to appear in AI Overviews or AI Mode, and that normal SEO best practices still apply.¹¹ So rewriting pages mainly to be "extractable" is not something Google asks for. A rewrite can change how a page ranks and converts in regular search, for better or worse. You only know if you measure both.

To be fair, citations are not worthless. Seer Interactive studied 3,119 informational search terms across 42 organizations. When a brand was cited in a Google AI Overview, its organic click-through rate was 0.70%, compared with 0.52% when it was not cited. That is about 35% more clicks. Paid click-through was about 91% higher.¹² Seer itself notes this is a correlation. Stronger brands may simply be more likely both to be cited and to get clicked.¹² The same study found organic click-through on queries with AI Overviews fell 61% between June 2024 and September 2025.¹²

So the honest summary is this. Being cited seems better than not being cited. But a citation is one input, not the outcome. The outcome is whether buyers choose you.

A measurement framework: track each signal separately

The fix is not to stop tracking citations. It is to stop treating any one number as the whole story. Here is a simple framework with eight signals, what each one tells you, and what each one misses.

Two practical notes on using this table:

• Run each prompt more than once. AI answers change from one run to the next. A single check can make a brand look recommended or ignored by chance. Our guide on how many checks you need for stable AI answer data covers this in detail.
• Read share of voice carefully. A share of voice number built only from mentions will count Oxlade's negative mention the same as Kestrelly's top pick. See how to read share of voice for how to avoid that trap.

How to test a change instead of chasing citation counts

Suppose your team wants to rewrite a set of pages to front-load key facts. Don't roll it out everywhere and watch the citation count. Run it as a test. Here's a simple way to do it.

1. Pick a group of similar pages. For example, 40 product comparison pages or 40 how-to guides. They should get similar traffic and serve similar intent.
2. Split them into two groups. One group gets the change (the "changed" group). The other stays as it is (the "control" group). Split them at random, or match them in pairs by traffic, so the two groups start out alike.
3. Write down your prompts and record a baseline. Choose the buyer prompts that relate to these pages. For a few weeks before the change, track mentions, citations, recommendations and position for those prompts, run several times each. At the same time, record organic clicks, impressions and conversions for both groups.
4. Make the change to the changed group only. Change one thing at a time if you can. If you rewrite the intro, add a table and change the title all at once, you won't know which one mattered.
5. Track AI and organic metrics together for the same period. Keep running the same prompts. Keep pulling Search Console and analytics data for both groups.
6. Compare the change, not the totals. Ask how much the changed group moved compared with how much the control group moved over the same weeks. If both groups gained 10% in citations, the change probably didn't cause it. Something else, like an AI model update, did.
7. Decide on business results. A change that raises citations but lowers recommendations, organic clicks or leads is not a win. A change that leaves citations flat but raises recommendations might be.

This kind of comparison is the same idea scientists use to separate cause from coincidence. It is slower than watching a dashboard go up. But it tells you what actually worked.

What to do this month

• Split your reporting. Report mentions, citations, recommendations and position as separate lines. Never add them into one score without also showing the parts.
• Find your ghost citations. List the prompts where your page is cited but your brand isn't named. Ask what the AI recommended instead, and why.
• Find your uncited wins. List the prompts where you are recommended without a link. Don't "fix" pages that are already doing their job.
• Look beyond your own site. If competitors get recommended with no citations, their strength probably lives in reviews, comparisons and discussions across the web. Work on how your brand is talked about there.
• Test before you roll out. Treat any "make it more extractable" rewrite as an experiment with a control group, and judge it on business results.

See where you're cited versus where you're actually recommended

Ripplix tracks mentions and citations as separate signals across the AI tools your buyers use, so you can see the prompts where your page supplies the facts but a competitor gets the recommendation. Start with a free report to see your own gap.

Get your free AI Visibility Report →