**TL;DR** — In April and the first two weeks of May 2026 we audited 24 client brands for unlinked brand mentions in Google AI Overview answer cards: the cases where the composer names a brand in the answer prose but the citation slot points to a third-party page rather than the brand's own site. Across 3,840 tracked queries the cohort brands were named in 612 answer cards; 47% of those mentions carried no citation back to the brand's domain. The composer linked instead to a category-comparison page (38% of the unlinked cases), an aggregator or review site (29%), a Wikipedia or Crunchbase entity page (18%), or a competitor's own page (15%). The pattern concentrated almost entirely on category and "best X for Y" queries and was nearly absent on branded queries. The shared root cause was structural: the brand's own pages framed the product as a product, not as one option inside a category, while the third-party pages did the comparative framing and the composer prefers the comparative framing for category answers. Two structural changes — publishing at least one first-party category-comparison page per major use case, and adding a one-sentence "we are the X for Y" positioning line to the product page — cut the unlinked-mention rate from 47% to 22% across a 30-day follow-up window.
Why we ran this audit
Last month's cross-engine citation work made it obvious that "share of citations" was an incomplete metric. A brand can win or lose share without anything happening on its own pages, because the citation slot is determined as much by how third-party pages frame the brand as by what the brand publishes itself. The follow-up question several clients asked, almost word-for-word, was: how often is our brand actually being named in the answer text without earning the citation slot? They wanted a number, not a vibe — and we did not have one. Anecdotally we saw it constantly; we had never measured it.
The second motivation was a budgeting problem. If you assume every brand mention is accompanied by a citation, you under-count exposure and over-count attributable traffic. The brand is getting impressions in the answer text that the citation log does not record, and the user — who has just read a comparative sentence naming three vendors — does not click any of the citations because the answer has already done the comparison job. We wanted to know how often this was happening so we could stop telling clients "citations are flat, so AI search is not moving the brand" when in fact the brand was being named in the answer prose at twice the citation rate.
How we ran the measurement
24 client brands — 9 SaaS, 7 DTC, 5 B2B services, 3 marketplaces — and for each brand a fixed basket of 160 queries split across branded, category and "best X for Y" intent. Each query was issued three times across the audit window from clean profiles on the same IP block to keep personalisation noise down; only answer cards that surfaced consistently across at least two runs were kept. For every answer card we logged two things: the full answer prose, parsed for any mention of the brand by name (including known aliases), and the cited URL set. A mention with a citation that resolved to the brand's domain was a "linked mention". A mention whose corresponding sentence carried no citation to the brand's domain — even if the slot pointed somewhere else — was an "unlinked mention". We did not count brand mentions inside cited paragraphs from third parties, because those are an entirely different mechanism we will write up separately.
Two normalisations matter for reading the numbers below. We treated the cited URL set per answer card as a whole — if any citation in the card pointed to the brand's domain, the brand was classified as linked for that card, even if the slot was not the one adjacent to the brand-mention sentence. That is a conservative choice; if anything, it understates the true unlinked rate. And we classified intent per query before the audit, not after — branded, category, or "best X for Y" — so the per-bucket numbers are clean rather than reverse-engineered from where the citations landed. The split mattered: aggregating across intent buckets would have hidden the most actionable pattern in the data.
The shape of the unlinked-mention pattern
Across the cohort, 47% of brand mentions in answer cards carried no citation to the brand's own domain. Inside that headline the per-intent split was: 8% on branded queries (essentially noise — when someone searches your brand name, the answer card almost always cites you), 41% on category queries, and 71% on "best X for Y" queries. The "best X for Y" number is the operationally important one: by the time a query reads "best CRM for early-stage SaaS", roughly three in four brand mentions in the answer card belong to a brand that did not earn the citation slot. The composer is doing the comparative work using a third-party page, and the brand is showing up as a name inside that comparison.
The destination of the unlinked-mention citations was concentrated and consistent. 38% pointed to first-party comparison pages on category-defining sites — G2, Capterra, software review hubs in SaaS; comparison and "best of" editorial in DTC; analyst sites in B2B services. 29% pointed to aggregators and review sites that scrape or curate vendor data without doing real first-party analysis. 18% pointed to Wikipedia or Crunchbase entity pages, which the composer trusts as a canonical short description of what the brand is. And 15% pointed to a competitor's own page — usually a "vs." or "alternatives to" page where the cohort brand was named as one of the alternatives. The last category was the most painful to look at, because the click that the brand mention generated went to a competitor's domain.
Driver one: comparative framing wins the citation slot
Reading the cited third-party pages, the consistent structural difference from the brand's own pages was framing. The brand's page said "Our product does X. It has features A, B, C. Here is pricing." The cited third-party page said "For X, the leading options are Brand A, Brand B, Brand C. Brand A is best for early-stage teams because of its lower starting price. Brand B is best for enterprise because of its admin features." The composer needs comparative framing to answer a comparative query, and the brand's own page does not supply it — so the brand is named in the answer text from the third-party paragraph, and the citation slot goes to the page that did the comparative work. The brand gets the mention; the third party gets the click.
The fix is unpleasant for product marketers but operationally simple. We helped seven of the 24 brands publish a first-party category-comparison page per major use case — "Best [category] for [use case]", written honestly and including competitors. The pages followed a tight template: a category definition paragraph, a 3–5 row comparison table, a paragraph per option naming the trade-off, and a "we recommend us when…" closing line that named the specific buyer profile we fit. After 30 days, queries where one of those pages existed saw the unlinked-mention rate fall from 71% to 28% on "best X for Y" intent — the brand started earning the citation slot for its own name in answer cards built off the new comparison page. The pages converted poorly relative to the rest of the site; their job was citation, not conversion.
Driver two: the missing one-sentence positioning line
The second pattern showed up on category queries (not "best X for Y", just "what is the best CRM"). Here the unlinked-mention rate was 41% and the cited pages were often the brand's own product page — but with the citation pointing to a competitor whose product page led with a one-sentence positioning claim like "Vendor X is the [category] for [audience]". The composer reads that sentence as a clean answer to "what is X" and lifts it. The brand whose product page led with feature copy or a hero benefit got named in the surrounding text but did not earn the slot. The cost of fixing it was one sentence added near the top of the product page; the lift in citation share over 30 days, on a per-page basis, was material.
The sentence has to follow a specific shape to work: "[Brand] is the [category] for [audience or use case]." Variants we tested — leading with a benefit, leading with a feature, leading with a quote — all underperformed. The composer is looking for a category-defining clause, not a brand-defining one, and the moment the sentence answers the category question ("what is the best CRM for early-stage SaaS") with the brand as the subject of the sentence, the citation slot becomes available. This is closer to old-school value-proposition copywriting than to anything specifically GEO, which is mildly deflating; the new layer just makes the old discipline pay differently.
Driver three: entity pages are doing the canonical-description job
The 18% of unlinked mentions that linked to Wikipedia or Crunchbase entity pages were a different problem. The composer wanted a one-sentence "what is this brand" answer, the brand's own pages did not have one in a clean enough form, and the entity page did. The lift here is not to publish a comparison page but to make sure the brand has a clean, scannable, schema-marked entity description on its own About page or homepage: the kind of paragraph a third party would lift verbatim to describe you. We saw two clients gain citation share on entity-style answers within 14 days of publishing a one-paragraph entity description with `Organization` schema and a `description` field that matched the prose. The composer prefers a first-party source if it exists in a usable form; the entity-page citations are mostly happening because no usable first-party form exists.
There is a Wikipedia subtlety worth flagging. Brands that already have a Wikipedia article cannot easily change the cited paragraph — Wikipedia's editorial standards are independent of the brand's preferences, as they should be. What you can do is make sure your own About page is at least as good a source for the same factual content, in a more recent and more verifiable form, so that on queries where the composer has a choice it prefers yours. We did not see any direct lever to influence which one it picks; we did see the choice shift over time on pages where the brand's own description got materially better.
What changed in our weekly reporting
We split the citation column into three: linked mentions (brand named, citation lands on us), unlinked mentions (brand named, citation lands on a third party), and non-mentions (brand not in the answer prose at all). The middle column is the new one. It is small in absolute terms — most queries do not name us at all — but it is the column where the work has the highest leverage, because the brand has already cleared the hardest bar (being part of the answer) and is losing only on framing. We also added a "third-party host" sub-column so we can see, week over week, which sites are winning citation slots that name us — that list is now the input to a small first-party-comparison-page roadmap rather than a "we should pitch them for a link" prospect list.
We dropped one report line. The previous monthly board had a "brand mentions across AI search" number that conflated linked and unlinked mentions into a single vanity figure. It was going up steadily for everyone in the cohort and was almost completely uncorrelated with anything we cared about. Replacing it with the split made the brand pages we needed to fix immediately obvious; replacing it without telling clients first made several of them think the number had dropped because we changed the methodology. The lesson — file under "boring" — is that when you change how a metric is computed, the same chart with the same scale will lie about itself unless you note the methodology change on the slide.
- 01Treat unlinked brand mentions as their own metric. In our cohort 47% of answer-card brand mentions carried no citation back to the brand's domain — that is the column where a structural fix actually moves the number.
- 02Publish first-party category-comparison pages for each major use case. After 30 days, the unlinked-mention rate on "best X for Y" intent fell from 71% to 28% on queries where one of those pages existed.
- 03Lead the product page with a one-sentence category clause: "[Brand] is the [category] for [audience]." Variants that led with benefits or features did not earn the category-query citation slot.
- 04Ship a clean, schema-marked entity description on the About page. The composer cites Wikipedia and Crunchbase when no usable first-party entity description exists — supply one and it often switches.
Where this argument breaks
For very low-awareness brands the audit is not yet diagnostic — the answer card does not name you regardless of framing, and the lever is brand familiarity (PR, partnerships, share of voice in the source corpus) rather than on-page structure. For Wikipedia-cited entity-page mentions on long-established brands, the composer's preference for a third-party canonical source is hard to dislodge inside a 30-day window; we would not promise the same lift there. The cohort skewed toward English-language Western markets; in Chinese-language AI search (元宝, 文心一言, Kimi) the third-party pages doing the comparative framing are different — 知乎 answers and 36kr columns dominate where G2 and Capterra do not exist — and the playbook for owning the comparative frame transfers in shape but not in target site. Our window was 60 days, which is short enough that we cannot tell whether the 47% → 22% lift persists at six months or whether the composer will rebalance back toward third-party comparisons as more brands publish their own. We will re-audit in October. The interim claim is narrower than it sounds: name and structure are the levers that exist right now, and they pay measurably more than another month of feature-led product copy.