“Coverage is up again this quarter.” If you’ve sat through a monthly PR review, you’ve heard that line, and you’ve probably said it.
Then, in the same month, you type “best tools for X” into ChatGPT and your name doesn’t come up.
Both things are true at once. The clip count is climbing. The AI doesn’t know you.
I used to think exposure just stacked — coins in a jar, and the denomination doesn’t much matter. Fifteen mentions is fifteen mentions.
But somebody went and sorted 410,000 of them by where on the page they actually sat, and the ones the models pick up have a shape. So let’s look at what that shape is, and at which column your coverage list is missing.
”Coverage Is Up. So Why Doesn’t The AI Know Us?”
The people who put data behind that question are BrandMentions, in an observational study published in late July 2026. The scale is what makes it worth your morning: 240 brands across six industries, about 410,000 public mentions, over the 90 days from 19 April to 17 July 2026.
They watched three AI surfaces — Google’s AI Overviews, Perplexity, and ChatGPT’s browsing mode — and asked them commercial-intent questions. “Best”, “compare”, “alternatives to”. The way you talk to a model when you’re lining up candidates before you buy.
Then they split the brands into the ones that came up over and over in those answers and the ones that basically never did, and compared how each group’s mentions were distributed. Not how many. How they sat. Start with the simplest question: where on the page is the mention?
For the brands that show up repeatedly in AI answers, about 70-80% of mentions sat in the body copy, a table, or a list. Mentions parked in a footer, a sidebar, or the boilerplate came to under 5%. Boilerplate is the stuff that’s identical on every page of a site — the company blurb in the footer, the row of customer logos in the sidebar. It’s the part a human eye slides straight past.
Turns out the mentions that land in AI answers mostly aren’t sitting there either. Which, for anyone who has spent email after email negotiating a spot in a logo strip, is an uncomfortable place to land. I’ve done that. The logo went in. On this data, I’d guess it did less for us than I assumed.
Who Wrote It Splits The Two Groups Cleanly
Next question: whose page was the mention on?
For the high-visibility brands, roughly 70-80% of public mentions came from third parties — somebody else’s article, not the brand’s own site or blog.
The brands that almost never came up looked different: owned media made up about 30-40% of their mentions, with third parties at 60-70%.
So building out your own content isn’t the mistake here. But the more of your pile comes from your own pages, the closer its shape gets to the low-visibility group’s.
The volume comparison — mentions versus backlinks — is the one I went through in the piece on why your backlinks barely register, where Ahrefs put brand mentions at a 0.664 correlation with AI visibility against 0.218 for backlinks (correlation just measures whether two things rise together — 0 is unrelated, 1 is perfect lockstep). This study adds the other axis: not how many mentions, but what shape they’re in.
The Biggest Pile Has No Link On It
Here’s the number that changes what you ask for in a placement conversation. Among third-party mentions around the high-visibility brands:
- Plain text, no link: about 50-60%
- Hyperlinked: about 25-35%
- Links to a marketplace or profile page: about 5-10%
The single largest group is an unlinked mention. Just your name, in a sentence, going nowhere.
So if you’re screening outlets on whether they’ll give you a link, you’re voluntarily skipping the biggest bucket. Links still do things — this isn’t an argument for ignoring them. But “no link, so nothing to report” is a scoring rule that doesn’t match what the models are reading.
What Sits Next To Your Name Decides Whether It Sticks
Placement was where; the last piece is what’s around it.
About 35-40% of the high-visibility mentions landed within 5-10 words of a category term. A category term is the phrase that says what kind of company you are — “project management tool”, “payroll software”. The bit a reader already knows without being told.
Another 20-25% sat near comparison language: “versus”, “alternative to”, “instead of”. And 10-15% were bare name-drops with no descriptive context at all.
That last slice is the trap. A human reading a bare name-drop fills in what you do from the rest of the article. The thing pulling names out of the page doesn’t fill anything in — if the sentence doesn’t say what you are, the mention teaches it nothing.
So the sentence you’re actually fishing for is a boring one: your category and your name in the same breath, ideally in a list of options. Not a paragraph about you. One clause that connects the two.
Read These As A Difference In Shape, Not A Target Number
Three caveats before you take any of this to a slide, and they matter. BrandMentions sells brand-mention monitoring, so “mentions drive AI visibility” is a conclusion the house wins on. Every figure is a range — 70-80%, under 5% — with nothing published on how likely those gaps are to be noise (no significance tests, no confidence intervals, which is the band a number would wobble inside if you ran the study again on a different sample). And it’s correlation: nobody showed that moving a mention into the body copy causes the model to name you.
So don’t nail 70% or “under 5%” into an internal target. Read it as the relative thing it is: the two groups’ mentions were distributed differently. Maybe I’m giving one vendor’s ranges more weight than they’ve earned. But if the distributions really do differ, then which one your own mentions resemble is something you can count starting today.
Add One Column To Your Coverage List
The mentions the AI picks up have a shape, and it’s specific. Somebody else’s article, in the body copy or a table, with a category term sitting next to the name. Whether there’s a link is the next question down, not the first one. And your coverage list can’t see any of it, because it counts clips.
So the first move is a sorting job you can do this afternoon. Open your last three hits and put each one in a bucket: name in the body text, table, or list — or name in the footer and the logo strip. That’s the column your report is missing, and a quarter’s worth of clips doesn’t take long to sort.
The second, and I’d hold this one lightly but do it anyway: put that split next to how often you actually turn up in AI answers, and look at both once a month in the same shape. A single reading tells you nothing about direction, so the point is asking the same questions repeatedly rather than asking better ones once.
We’ve all been counting the clips. Almost nobody’s been counting where they landed on the page. Which means there’s still a fairly cheap edge sitting in a column nobody’s added yet.
Sources
- BrandMentions (Cornelia Cozmiuc), “Do Brand Mentions Really Influence AI Visibility? The 2026 Evidence, Explained”, July 2026, brandmentions.com (about 410,000 public brand mentions tied to 240 brands across six industries; 90 days from 19 April to 17 July 2026; commercial-intent query answers observed across Google AI Overviews, Perplexity, and ChatGPT browsing mode. Brands appearing repeatedly in AI answers drew about 70-80% of public mentions from third-party sources; brands appearing almost never had owned media at about 30-40%. Third-party mentions around high-visibility brands broke down as plain text about 50-60% / hyperlinked about 25-35% / marketplace or profile links about 5-10%. Mention placement: body copy, tables, and lists about 70-80%, footer, sidebar, and boilerplate under 5%. About 35-40% of high-visibility mentions fell within 5-10 words of a category term, about 20-25% near comparison language, about 10-15% were bare name-drops with no descriptive context. External data referenced in the article: Ahrefs put AI-visibility correlation at 0.664 for brand mentions against 0.218 for backlinks; AirOps found 85% of brand mentions came from external domains. Limits: the publisher sells brand-mention monitoring tooling and has a commercial interest in this conclusion; figures are ranges with no significance tests or confidence intervals, and the study is an observation of correlation, not proof of causation)
