Google cited your page and recommended your competitor. That happened 69% of the time

Most AI visibility reporting counts one thing: how often your domain appears as a cited source. That number is easy to collect, it goes up when you do the obvious work, and it can rise while you lose the customer. A recent analysis of Google AI Overviews put a figure on the gap, and the figure is large enough to change what you publish.

What the analysis found

The study, conducted by Lily Ray and reported by Search Engine Land, looked at 100 B2B queries of the form "best [category] software" in Google AI Overviews, collected at three checkpoints in 2026: 15 April, 15 May and 8 June. Answer text and cited sources were gathered using Ahrefs Brand Radar. Two outcomes were measured separately for each query: whether a brand's own self-promotional listicle was cited as a source, and whether that brand was recommended in the answer.

Eighty of the 100 queries returned an AI Overview. Across those, self-promotional listicles were cited 323 times. In 224 of those citations, roughly 69%, the AI Overview used the brand's own page as a source and then recommended somebody else, frequently a better-known competitor named inside the brand's own article.

Two caveats before drawing conclusions. This is one query type in one vertical, so the specific 69% should not be read as a universal rate. And AI answers are unstable: independent research has repeatedly found that the same prompt produces different brand recommendations across runs, which means any single measurement is a sample rather than a fact. What survives both caveats is the structural finding, which is the useful part: a citation and a recommendation are produced by different things, and only one of them wins the customer.

Why the self-promotional listicle backfires specifically

The mechanism is not mysterious once you look at what these pages contain. A "best [category] software" article written by a vendor in that category does three things at once. It demonstrates topical relevance, which earns the citation. It names competitors, because a listicle with one entry is not a listicle, which supplies candidate answers. And it ranks itself first with reasoning that an evaluating system has no way to verify, because the claim is self-interested by construction.

So the page is an excellent source and a poor recommendation. It reads as a comparison document, and comparison documents are exactly what a model wants when assembling an answer to a comparison query. The brands named inside it arrive with third-party corroboration attached; the brand that published it arrives with its own endorsement, which carries almost no weight. You have written a well-structured summary of your competitive set and handed it over.

This is the same asymmetry we found in the earned media data: AI search cites earned media far more often than brand-owned content. The new finding sharpens it. It is not only that owned content is cited less. It is that a particular kind of owned content is cited a lot and converts that citation into a recommendation for somebody else.

Which content earns citation, which earns recommendation

If the two outcomes come from different properties, they can be planned for separately. Roughly, citation rewards a page that is easy to extract from, and recommendation rewards a brand that is easy to corroborate. Very little content does both well, and knowing which one a given format buys you is the practical use of the whole finding.

The row that matters most for small and mid-size businesses is the fifth one. Recommendation is substantially earned off your own site, which is uncomfortable if your content plan is a publishing calendar. It means review profiles, directory listings, being named in other people's comparisons and appearing on podcasts are not brand-awareness extras; they are the input to the outcome you actually want.

What to publish instead, in priority order

These are ordered by how much recommendation each buys per unit of effort, which is a different order from how easy they are.

1. One number only you have. Aggregate something you already collect: quote-to-close rates, average project timelines, the distribution of a metric across your customers, the failure modes you see most. A single well-defined statistic with a stated method and date earns citations for years and carries your name every time.
2. A named decision framework. If your sales conversations always resolve to the same three or four criteria, write those down as a framework with thresholds. Frameworks get recommended because the recommendation and the citation are the same thing: there is no way to use it without naming it.
3. Specifics about your own offer that a model cannot infer. Real pricing structures, what is and is not in scope, turnaround times, the conditions under which you decline work. This is the content that answers "who should I use" for the narrow query where you are genuinely the right answer.
4. Third-party presence, treated as a deliverable. Review platforms and directories relevant to your category, expert roundups, podcasts, and being named in other people's comparisons. Assign it to someone with a target, the same way you would a publishing target.
5. Comparisons where you are not a participant. Comparing two tools you do not sell is credible in a way that comparing yourself to a rival is not, and it earns the citation without handing over the recommendation.

Notice what is absent: more volume. Adding another guide to a category already saturated with guides raises your citation count at best and does nothing for recommendation, because a recommendation needs a reason that is specific to you. If you are building pages to convert the traffic that arrives from these systems, the mechanics are different again and we covered them in landing pages for ChatGPT and Perplexity traffic.

Fix your measurement, or you will keep optimising the wrong number

Most AEO dashboards report citations, mentions and share of voice. All three can rise while your recommendation share falls, and the listicle finding is the clean example of how. Three changes make the reporting honest.

1. Record citation and recommendation as two separate fields. For each tracked prompt, log whether you were cited as a source and, independently, whether you were named as a recommended option. The gap between those two columns is the number worth managing. Ray's analysis is only possible because it measured them separately.
2. Sample repeatedly, and report a rate rather than a state. Because AI answers vary between runs on identical prompts, a single check tells you very little. Run each prompt on a fixed schedule and report the share of runs in which you were recommended. Treat a change smaller than your run-to-run variation as noise.
3. Tie it to revenue, not to visibility. Segment AI referral traffic in analytics and track its conversion rate and value separately. AI-referred visits arrive with more context than a cold organic click and the reported conversion gaps are large, which means a small traffic number can carry a disproportionate share of pipeline, and is worth valuing correctly.

The measurement discipline matters more here than in traditional SEO because the feedback is so much noisier. A ranking position is stable enough to act on from one observation. A recommendation is not. Our AEO and GEO guide covers the page-level formatting that makes content extractable, and what llms.txt does and does not do deals with a related measurement myth.

What this does not mean

It does not mean citations are worthless. A cited page still gets a link card and still gets clicks, and citation is a prerequisite for some recommendations rather than a substitute for them. It also does not mean you should stop writing about your own category; it means you should stop ranking yourself inside your own comparison and expecting that to count as evidence.

It also does not generalise cleanly to every query type. The analysis covers B2B software comparison queries, which are unusually crowded with vendor-written listicles. A local service query, a how-to query and a product query each behave differently, and the honest position is that the structural lesson transfers while the 69% does not. If your category is not saturated with vendor listicles, your exposure is smaller, though the underlying point still holds: a system assembling an answer will always prefer a claim it can verify to a claim you make about yourself. Building the kind of content that survives that test is most of what content work is for now.

Frequently asked questions

What is the difference between an AI citation and an AI recommendation?

A citation means an AI answer used your page as a source, usually shown as a link card alongside the answer. A recommendation means the answer named your brand as an option the user should consider. They are separate outcomes: a page can be cited as evidence while the answer recommends a competitor. Citation rewards content that is easy to extract from; recommendation rewards a brand whose merits can be corroborated independently. Measuring only citations overstates your position.

How often does Google cite a brand without recommending it?

In an analysis of 100 B2B "best [category] software" queries by Lily Ray, collected on 15 April, 15 May and 8 June 2026 using Ahrefs Brand Radar, 80 queries returned an AI Overview and self-promotional listicles were cited 323 times. In 224 of those cases, about 69%, Google cited the brand's own page but recommended a different brand. That rate is specific to vendor-written comparison content in B2B software and should not be read as a universal figure, but the structural pattern it shows is the point.

Should I stop writing "best [category]" listicles about my own category?

Stop ranking yourself inside them. The format earns citations well, but when you place yourself first with reasoning that cannot be verified, and name competitors as the other entries, you supply the candidates and the evidence while your own claim carries little weight. Keep the page, remove your own entry from the ranked positions, disclose your interest, and turn it into a selection framework with criteria a reader can apply. Comparing tools you do not sell is credible in a way that comparing yourself to a rival is not.

What content actually gets recommended by AI answer engines?

Content whose claims cannot be separated from you, plus independent corroboration. In practice: one original statistic you produced, with a stated method and date; a named decision framework, because using it requires naming it; specific, verifiable detail about your own offer such as pricing structure, scope and turnaround; and third-party presence, meaning review platforms, directories, expert roundups and being named in other people's comparisons. That last category is earned off your own site, which is why recommendation cannot be solved by a publishing calendar alone.

How should I measure AI search visibility properly?

Log two separate fields per tracked prompt: whether you were cited as a source, and whether you were recommended as an option. Report the gap between them. Because AI answers vary between runs on the same prompt, run each prompt on a fixed schedule and report the share of runs in which you were recommended rather than a single snapshot, and treat movements smaller than your run-to-run variation as noise. Then segment AI referral traffic in analytics and track its conversion rate and value separately, because a small traffic number can carry a disproportionate share of pipeline.

Are AI citations still worth anything if they do not lead to recommendations?

Yes. A cited page normally gets a link card and receives clicks, and for some query types citation is a step towards recommendation rather than an alternative to it. The problem is treating citation count as the goal, because it can rise while your recommendation share falls. Value citations as a traffic and credibility signal, and manage recommendation share as the commercial outcome.

The takeaway

Citation and recommendation are two different outcomes produced by two different things, and only one of them wins the customer. A listicle about your own category is superb at getting cited and poor at getting chosen, because it supplies candidates and self-endorsement at the same time. Publish the things that cannot be detached from you, one original number, a named framework, verifiable specifics about your own offer, and treat third-party presence as a deliverable rather than a bonus. Then log citation and recommendation as separate fields, sample repeatedly because the answers move, and judge the work on recommendation share and revenue rather than on how often your domain appears.