In October 2025, Yext published one of the largest public studies of AI citations anyone had run to that point. Its analysts pulled 6.8 million citations across 1.6 million answers from ChatGPT, Gemini, and Perplexity and sorted every one by where it came from.

The headline number traveled fast: 86% of citations came from sources a brand already controls. First-party websites alone accounted for roughly 43 to 44%, the single largest source category.

That number got quoted in a hundred LinkedIn posts. The one underneath it did not. Gemini pulled 52.1% of its citations straight from brand-owned domains, while ChatGPT pulled almost half of its citations from third-party listings instead. Same query, same brand, two engines that disagreed about whose page deserved to be named.

So a company can be the source behind an answer and still be invisible inside it. The engine used your page. The reader never saw your name. That gap, between being used and being named, is what this issue is about.

This briefing covers one signal, one build, and one move. It takes about 12 minutes to read, and you can run a 40-minute exercise before next Thursday.

When a Google AI Overview renders, the click-through rate to the regular blue links below it drops between 30% and 60%, depending on the query type, with informational and definitional searches being hit the hardest. That range comes from a study of 1,000 AI overviews that DigitalApplied ran across 30 verticals in April 2026. What replaces the lost clicks is the small set of attributed sources the model names inside its answer, typically three to five per response. As that study puts it, for a large class of queries, you are either in that citation set or your visibility for the search drops sharply, even if you still rank below the Overview.

So the citation slot may become the main traffic a page can still earn from that result. That changes the math on your content budget, and it raises a second question the 86% headline hides: not all citation slots are equal.

The most direct evidence on that comes from outside the marketing-vendor world. Pew Research Center tracked the real browsing of 900 US adults across roughly 69,000 Google searches in March 2025. When an AI overview appeared, users clicked a source link inside the summary in just 1% of visits. For comparison, they clicked a traditional result link 8% of the time when a summary was present, and 15% when one was not. The sources parked in the panel, the "see also" layer, are close to a dead-click surface. (Pew measured Google AI Overviews specifically, and Google has disputed the methodology; treat it as one strong signal, not the last word.)

That 1% is the number to sit with. It means the value of a citation depends heavily on where it lands. A source woven into the sentence a reader is actually reading is in the path of attention. A source relegated to a tray that the reader has to open mostly is not. An older measure points the same way: Authoritas found in early 2025 that pages cited in Perplexity answers earned 2.3 times more referral clicks than pages merely mentioned by name. That figure is over a year old, and the engines have shifted under it, so treat it as directional. Both readings land in the same place: placement, not mere presence, is what turns a citation into a visitor.

What earns a better placement is structure, and here the citation studies are consistent. DigitalApplied found that pages carrying their own named-source citations in the body were cited 2.1 times more often, and pages with clean schema markup 2.3 times more often, than otherwise comparable pages. The engines reward content they can lift a clean, self-contained, attributable claim from.

For a revenue leader, the decision this forces is narrow and concrete. Stop measuring whether AI engines mention your company. Start measuring where. A brand-mention dashboard that counts every appearance equally is averaging your most valuable placements together with your least valuable ones and reporting the blend as progress. And the stakes are not just traffic: if your company is absent from the decision-stage answers buyers read first, you can lose them before they ever reach your site.

Yext is a publicly traded brand-visibility company, NYSE-listed, that sells software for keeping business information consistent everywhere it appears online. In 2025, it had a measurement problem of its own: its customers kept raising concerns. AI engines were citing brands, but nobody could say which kind of content earned the citation or why two engines answering the same question cited different sources.

Most of the market was guessing. The loudest theory held that AI answers ran on Reddit and user forums, so brands should pour effort into community posts. Yext decided to measure it instead of arguing about it.

The study ran across 1.6 million queries between July and August 2025, classified 6.8 million resulting citations by source type, and broke the results down by engine, by industry, and by whether the question was branded or generic. The scale is what makes it useful. A 50-query sample tells you about 50 queries. A 6.8-million-citation sample tells you about the system.

Two findings cut against the prevailing advice. Forums like Reddit accounted for just 2% of citations once the analysts controlled for location and query intent, a fraction of the weight the market assumed. And the engines diverged sharply on what they trusted: Gemini behaved like a strict search engine, taking 52.1% of its citations from structured brand-owned pages, while ChatGPT leaned on third-party listings for nearly half of its citations. For unbranded, generic questions, the kind a buyer asks before they know your name, first-party websites and local pages supplied close to 60% of all citations.

The lesson Yext drew from its own data is the one that matters for everyone else. Visibility in AI answers is not random, and it is not mostly earned on forums you don't own. It is earned on structured, consistent, machine-readable content on pages you do control. The brands most likely to be cited are the ones whose pages make a claim easy to extract, verify, and attribute, the same trait that, on the engines that place sources inline, tends to earn the inline slot rather than the buried one.

The catch worth naming honestly: this is Yext's own research, and Yext sells a product built on the conclusion. The 6.8-million-citation scale and the engine-level breakdowns are verifiable and have been widely reported. The interpretation deserves the same skepticism you would bring to any vendor study. The numbers stand on their own; weigh the framing for yourself.

Run this in about 40 minutes. It needs nothing but the AI tools your buyers already use.

Step one. Write down your eight most commercially valuable buyer queries, in the words a buyer would actually type, not your marketing language. Mix three research-stage questions ("What is X and why does a company need it?"), three comparison-stage ("Best X for mid-market B2B" and "[competitor] alternatives"), and two decision-stage ("Is X worth it for a Series B company?").

Step two. Run each query in all four engines: ChatGPT, Perplexity, Gemini, and Claude. For every answer where your company appears, mark one of two things: were you named inline, inside the sentence the reader actually reads, or were you relegated to a "sources" or "see also" list they would have to expand? The engines surface this differently, so know what you're looking at: Perplexity numbers its sources inline; Claude links inline when it runs a live search; Gemini and Google's AI Overviews often show sources below, beside, or inside a panel, which Pew's data suggests gets clicked far less than an inline link. Note which competitors landed inline where you did not.

Step three. Sort your results into three piles. Inline citations are slots you own. Source-list-only mentions are the highest-value fixes you have, because the engine already trusts your page enough to use it; it just isn't naming you where readers look. Queries where a competitor is inline, and you are absent, are where you are losing a fight that is actively being fought. Sort by query stage too: a gap on a decision-stage query, the one a buyer asks right before they choose, costs more than a gap on an early research question.

Step four. Act on the sort. Each pile points to a different fix, and the order matters more than the volume.

The single most useful output of this exercise is the second row: the pages an engine already trusts enough to use but not enough to name. Those are the cheapest wins you have, and a count-everything dashboard will never surface them.

When you want this measured properly, across 50 prompts, four engines, and three runs each, rather than a one-pass spot check, that is what our Citation Audit Method does.

Elizabeta Kuzevska Co-Founder, Revenue Experts AI https://revenueexperts.ai

If you've never checked whether AI engines can even read your site, run the free 60-second AI Visibility Audit first. It's the readiness layer.

If you already know your site is technically sound and you want the actual citation map that surfaces your category, competitors, or nobody across all four engines, the $497 AI Visibility Audit is the one. Five to seven days, 600 measured calls, a per-prompt diagnosis, and a ranked fix list.

For the method behind both, the Citation Audit Method pillar lays out the full framework.

Sources