Your brand gets cited inside a Google AI Overview. That feels like a win. New tracking from Lily Ray, VP of SEO & AI Search at Amsive, says it often isn't one. A citation and a recommendation are two different events, and AI Overviews are increasingly splitting them apart. Ray tracked 100 B2B "best [category]" queries across three checkpoints between April and June 2026. The pattern she found: brands got their own self-promotional "best" listicle cited by name, then watched the AI Overview turn around and recommend a competitor instead — often the exact companies their own listicle was written to beat.
Citation vs. recommendation: the split AI Overviews are making
Most generative engine optimization advice right now reduces to one instruction: get cited. Publish a "best CRM for X" or "top 10 project management tools" page, get it indexed, get it pulled into an AI Overview, declare victory. Ray's tracking breaks that shortcut. A citation tells an AI engine that a page exists and addresses the topic. It does not tell the engine that the page's opinion is the one it should repeat back to the user. Google's AI Overview builds its actual recommendation the way it builds any synthesized answer: by weighing multiple sources against each other, not by parroting whichever self-ranked list happened to get linked. Your "best X" page can be the citation and still lose the argument inside the very answer it appears in.
The method here matters, because this isn't a hunch. Ray ran 100 B2B "best [category]" queries through Google AI Overviews at three separate checkpoints between April and June 2026, pulling citation data through Ahrefs Brand Radar. She published the analysis on her Substack on June 17, 2026, titled Why Calling Yourself the 'Best' Could Be Helping Your Competitors Win in AI Search, and Search Engine Land covered the findings the next day, on June 18, 2026. Three checkpoints across three months is enough to rule out a one-off fluke. This is a pattern that held across a quarter of B2B search behavior, not a screenshot from one bad afternoon.
The data: cited 323 times, recommended a competitor 69% of the time
Across the queries Ray tracked, self-promotional "best" listicles — pages where a company ranks itself against competitors on its own site — got cited 323 times inside 80 AI Overviews. That's already a healthy citation count for anyone running a GEO citation strategy the standard way: publish the comparison, get the AI engine to link it, move on. Ray kept watching past the citation, into what the AI Overview actually told the user to buy. In 224 of those 323 citations — 69% — the brand whose own listicle got cited was excluded from the recommendation itself. The AI Overview cited the page, then recommended someone else, often naming the exact competitors that page was written to beat.
Ray's clearest example: a learning-management-system company built a "best LMS platforms" listicle and got it cited inside an AI Overview — and the tools the AI Overview actually recommended to the user were Kajabi, Thinkific, LearnWorlds, and Teachable. The company's own page, Oasis LMS's listicle, supplied the citation. It did not supply the answer. The AI Overview used the page as evidence the category and its leading options existed, then made its own separate judgment about which of those options to put in front of the user — and that judgment didn't include the company that built the page.
| SCENARIO | WHAT THE AI OVERVIEW CITED | WHAT THE AI OVERVIEW RECOMMENDED |
|---|---|---|
| Aggregate across 80 AI Overviews | Brand's own "best X" listicle (323 citations) | A competitor, not the citing brand, in 69% of those citations |
| The LMS example | Oasis LMS's "best LMS platforms" listicle | Kajabi, Thinkific, LearnWorlds, Teachable |
| All 100 tracked prompts | Self-ranked comparison content, present in the answer | A different vendor recommended in 74% of prompts |
“A citation is not a recommendation.”
Why AI Overviews cite you and recommend someone else
The mechanism isn't mysterious once you separate the two jobs an AI Overview is doing. Retrieval finds pages that plausibly answer the query — your "best X" listicle qualifies, because it's structured, on-topic, and easy to parse. That's the citation. Recommendation is a second, independent step: the AI Overview weighs what multiple sources say about the category, not just what one self-interested page claims, and it favors the version of "best" that shows up across independent sources over the version one vendor asserts about itself. A listicle where the publishing company ranks itself first is, structurally, the least trustworthy source in the room for the recommendation step, even while it's a perfectly good source for the citation step. The two jobs use the same page differently, and most GEO citation strategy right now only optimizes for the first one.
This lines up with what we've found in our own citation research. In our breakdown of what actually gets cited across engines, comparison content accounted for 32.5% of the citations we tracked — more than any other format. That was always a citation number, describing what gets pulled into an answer as a source. It was never a claim that self-ranked comparisons win the recommendation too. Ray's data supplies the piece that number always needed: getting cited as a comparison source and getting recommended as the best option are different outcomes, and a page can nail the first while losing the second to a competitor named inside its own text.
Think about it from the AI Overview's side for a second. It has already decided your listicle is worth citing, which means the retrieval step went fine — the page is indexed, structured, and topically relevant enough to surface. The recommendation step is where the engine has to decide whom to trust, and trusting the page that ranked itself first, on the exact question of who's best, is a strange thing for a synthesis system to do by default. It's the equivalent of asking a company to grade its own exam and expecting the answer key to still hold up against three other graders. Some AI Overviews clearly do exactly that cross-check, which is why the recommendation splits away from the citation in 74% of the prompts Ray tracked.
Ranking still drives AI Overviews recommendations — citation alone doesn't
Ray's data comes with a warning attached, not just a diagnosis. In an earlier interview with ppc.land, published May 14, 2026, she flagged that AI-search "hacks" — including self-serving listicles built purely to chase a citation — get detected and penalized once they're deployed at scale, because "hundreds of thousands of these pages" are being built industry-wide right now, and search engines notice patterns like that. She also flagged real legal exposure: unsubstantiated claims of competitor inferiority inside a company's own "best X" content can draw FTC attention, not just an algorithmic penalty. And she made the point that still holds underneath all of it: "it's almost always ranking number one on Google gets you more citations." Organic ranking, not citation-chasing tactics, remains the biggest lever behind who gets cited in the first place.
That warning isn't theoretical. Ray also found that sites leaning too heavily on self-ranked "best" pages started seeing organic ranking declines around January 20, 2026, and that the drop worsened through Google's May 2026 core update. Read those two data points together and the strategy implication gets uncomfortable: the same self-promotional listicle format that fails to convert citations into recommendations 69% of the time is also a format Google appears to be quietly discounting in organic rankings. You can lose on both ends at once — a citation that doesn't convert to a recommendation, sitting on a page that's sliding on the on-page ranking signals it depends on.
Put the two data points from Ray's reporting side by side and a lazy GEO citation strategy looks worse than doing nothing at all. A brand that never bothered with a self-ranked "best X" page isn't carrying organic-ranking risk from the format, and isn't handing an AI Overview a built-in list of competitor names to recommend instead. A brand that built the page, chased the citation, and skipped the harder work of earning it honestly is exposed on both fronts: a possible ranking penalty on one side, and a documented 69% chance the citation just advertises the field to a competitor on the other. That's not an argument for giving up on comparison content. It's an argument for building the kind of comparison content that survives scrutiny from both Google's ranking systems and an AI Overview's cross-referencing step, rather than the kind built purely to farm a citation count.
What to do: track recommendation rate, not just citation rate
Citation rate is the metric most GEO citation strategy dashboards report, because it's the easiest thing to detect: did the engine link my page, yes or no. Ray's data says that number, on its own, tells you less than it feels like it tells you. A rising citation count on your own "best X" page can sit next to a flat or falling recommendation rate, and most teams have no way to see the gap, because they never built a way to track the second number separately from the first. Our mention-rate reporting breakdown covers the same failure mode from a different angle: the dashboard tells you what got mentioned, not what got recommended.
None of this means stop building comparison content. It means stop treating a citation on your own listicle as the win condition. Ray's data is blunt about the gap: 69% of the time a brand's self-promotional page got cited, a competitor got recommended instead, and 74% of all tracked prompts showed some version of that split. Start measuring recommendation rate as its own number this quarter, or keep finding out about the gap the way most of these brands did — after a competitor's name showed up in an answer their own page was supposed to win.
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Josh leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.