82% of Your AI Citations Point Somewhere You Don't Control
For about eighteen months, the standard response to AI search has been to work on your website. Cleaner headings. More schema. FAQ blocks. Entity pages. An llms.txt file. Restructure the content so a model can extract an answer from it.
That work is not wrong. It is just aimed at a minority of the problem, and a run of research published in the last two weeks makes the size of that minority uncomfortably specific.
Writing in Search Engine Land on July 28, Jen Cornwell, Tinuiti's senior director of AI SEO innovation, reported that in the agency's Q1 2026 AI Citation Trends analysis, roughly 82% of AI citations pointed to earned media, not a brand's own site.
If that figure is even approximately right, then the dominant tactical response to AI search has been optimizing the smaller share of the surface, and most GEO budgets are allocated against the wrong constraint.
Source: Search Engine Land, Why creator content belongs in your AI search strategy (July 28, 2026)
The finding that makes it hard to argue with
A single vendor statistic deserves the scepticism we would apply to any other. What makes this one difficult to dismiss is that a completely separate study, using a completely different method, landed on the same conclusion from the opposite direction.
Victorious tested 175 brands across five verticals on eight AI platforms for its Q2 2026 Quarterly Search Report, covered by Dataconomy on July 29. Split across two cohorts of 140 and 150 brands, it reports two numbers that belong together:
- 96% recognition accuracy when engines were asked about a brand directly.
- 89% of brands never surfaced in answers to category research questions.
Ask an AI engine about your company and it will almost certainly get you right. Ask it who the good options in your category are and it will almost certainly not mention you.
That pairing kills a comfortable assumption that has underpinned a lot of AI-visibility work: the theory that if the model knows what you do, getting recommended is a matter of tuning. It is not. The model knowing what you do is nearly universal and nearly worthless on its own. Recognition is table stakes that 96% of brands already pass.
This has an immediate diagnostic consequence. The standard AI-visibility check — asking an engine "what is Acme?" and reading the answer — is testing the thing almost everyone passes. The prompt that actually discriminates is the category question, where 89% fail.
If your audit is built on brand-name prompts, it is measuring the wrong thing and will report good news indefinitely.
Source: Dataconomy, AI Recognizes Brands But Rarely Mentions Them In Category Searches (July 29, 2026)
The 2,000-page threshold
The Victorious data also points at what does move the needle, and it is not a property of your website.
The reported correlations are with volume of external coverage: 0.49 with referring domains and 0.45 with third-party mentions. Moderate, not decisive — but consistent in direction.
The starker figure is a threshold effect. Brands with fewer than 2,000 indexed pages mentioning them appeared in answers just 3% of the time.
Treat the exact number carefully; it comes from a 175-brand vendor sample and the verticals are not characterized as representative. But if a threshold of roughly that shape holds at all, the operational implication is significant: for a large share of mid-market brands, the binding constraint on AI visibility is not content quality and not technical SEO. It is the raw quantity of independent web pages that mention the brand at all.
That is a public relations and partnerships problem with a multi-quarter lag, not a sprint item. It does not fit neatly into a content calendar, and it cannot be fixed by a schema deployment.
Victorious additionally reports that 99.99% of citations at the comparison stage pointed to third-party sites. A figure that round at four nines almost certainly reflects a narrow query set rather than a law of nature, and should be read as "essentially all" rather than as a measurement. But the direction matches what anyone testing comparison prompts has observed: when a user asks which product is best, engines reach for review sites, forums and listicles, not vendor pages.
The practitioners agree, and they are more specific
Digiday published an unusually candid piece on July 27 in which agency leads described what actually works. Nate King of RPA, Heather Physioc of VML and David Dweck of Go Fish Digital told Kimeko McCoy that llms.txt files have shown little impact, that backlink schemes and keyword stuffing do not work in this regime either, and that in a zero-click environment success has to be measured through share of voice and brand mentions across assistants rather than click-through rate.
Their framing: AI visibility is an evolution of SEO, with roughly 80 to 90% of tactics carrying over and the remaining 10 to 20% being new and decisive.
The 80% is not in dispute and should not be surprising. Assistants read the web through crawlers. A page that is inaccessible, contradictory, thin or unattributed is a bad source for a summarizer for exactly the reasons it was a bad result for a ranker. Crawlability, structure, accuracy and clear entity signals are table stakes in both regimes.
What is worth noticing is why so much of the GEO market has concentrated on that 80%. It is not because that is where the leverage is. It is because it is what agencies already know how to sell and what software can already measure. Repackaging a technical audit as an AI-readiness audit is a low-risk product decision.
It is also, on this evidence, not where the outcome is decided.
Source: Digiday, reporting by Kimeko McCoy (July 27, 2026)
Where the citations actually go
"Earned media" is doing heavy lifting as a category. It spans journalism, forums, review sites, video and social — surfaces with wildly different acquisition costs and wildly different levels of brand influence. An aggregate 82% tells you where citations land, not what to do about it.
The category-level data is more actionable, and it says the answer is different for every industry.
Similarweb's 2026 Generative AI Landscape data, reported by Search Engine Journal, found ChatGPT citing sources in 22.6% of travel answers against 4.8% of education answers, on a 6.8% average across all answers. In travel, citations skewed 54.1% toward reviews and user-generated content.
Tinuiti's numbers show the same variance in the social channel: social content drove 13% of AI citations for apparel against 3% for over-the-counter health in Q2 2026. The spread is intuitive — regulated categories get answered from institutional sources, discretionary ones from peer opinion.
Less intuitive, and more important as a warning: Perplexity's social citation share fell from 31% to 13% in a single quarter.
That is not a category property drifting. That is an engine changing a product decision. Any off-site strategy built on a single engine's current source mix is built on something that can be halved in ninety days without notice.
The video finding that breaks the usual playbook
One data point deserves separating out because it inverts how most creator and influencer programs are constructed.
Tinuiti reports YouTube's inclusion in Google AI Overviews growing from 3.6 million to 36.2 million keywords year over year — a tenfold expansion. Citing OtterlyAI data, it adds that long-form video accounts for 94% of AI video citations, and that 40.83% of cited videos have fewer than 1,000 views.
Four in ten cited videos have under a thousand views.
Citation selection in this surface is substantially decoupled from audience size. The systems are picking video for what it says and how retrievable it is, not for how popular it is.
There is a corroborating anecdote from practice. Deloitte Digital's Jenny Kelly told Net Influencer on July 27 about a client case in which a YouTube creator with roughly 1,200 followers produced more LLM visibility than far larger accounts, because the content was structured for how models parse and surface information.
If that pattern holds, the economics of creator partnerships for AI visibility are genuinely different from the economics of creator partnerships for reach. You are not buying an audience. You are buying a well-structured, topically specific, independently authored document about your category — and a micro-creator can produce that as well as a large one, at a fraction of the cost.
You cannot buy your way past this
The obvious objection at this point is that if earned placement is slow, paid placement is fast. Google sells ads inside AI Mode. Buy the slot.
SE Ranking tested exactly that, in an analysis of 50,032 commercial keywords surfaced in B2The7's July 19 weekly brief. Ads appeared on 29.45% of those keywords. The question was whether advertisers running ads in AI Mode also showed up among the sources the generated answer cited.
They mostly did not:
- In 88% of ad keywords, the advertiser's domain was not cited at all for the term it had bid on.
- Only 1.95% matched at the exact-URL level.
Paid placement and earned citation are two almost entirely separate systems. A brand can pay to sit in the ad slot and still be absent from the list of sources the model tells the user it relied on.
This matters more than it first appears, because of how the two are read. An ad is understood by the user as an advertisement. A citation is understood as the model's evidence — the thing it consulted before forming a recommendation. Those carry very different weight with a buyer who is asking an assistant what to choose, and the second one is not available in the auction.
The usual caveats apply: these are SE Ranking's own figures from a single snapshot, and AI Mode's behavior is changing quickly. But the direction is consistent with everything else in this piece. Models cite what they judge authoritative, and that judgment is not for sale.
Source: B2The7, 5 Marketing & Digital Trends: Week of July 20, 2026 (July 19, 2026)
The uncomfortable reframe: this is a PR problem in technical costume
Two frameworks published in the same week arrive at this from different ends, which is usually a sign the underlying observation is real.
Rich Sanger's six-dimension "AI entity footprint" audit, published in Search Engine Land on July 27, formalizes the shift. Its dimensions are identity, differentiation, evidence, consistency, relationships and specialization — and crucially, the unit of analysis is not the page. It is whether a business's site, reviews, profiles and third-party mentions collectively produce one coherent entity understanding.
Kaleigh Moore's "Creator AEO" argument, covered by Net Influencer on July 28, comes from the opposite direction and lands in the same place. Her formulation: "the further that a source is from brand control, the more trustworthy it is." Her own writing puts a brand's own site at only 5 to 10% of the sources an assistant draws on — which frames owned content as insufficient rather than useless.
Put all of it together and the decisive variable is not the quality of your page. It is whether the sources an assistant prefers to cite in your category happen to say something about you, and whether what they say is consistent.
That is a public relations, community, review-management and creator problem wearing a technical costume. It sits in a different budget line and usually a different team from the one currently assigned to AI visibility, which is why it keeps not getting done.
What to actually do about it
Sorted by how long it takes to show up, because the honest answer is that the highest-leverage work here is also the slowest.
This quarter
Re-point your audit at category prompts. If 96% of brands pass the brand-name test and 89% fail the category test, then any measurement built on brand-name prompts is generating false comfort. Rebuild the prompt set around the questions a buyer asks before they know your name.
Inventory where you are already cited, by source type. Not a count — a breakdown. Journalism, review platforms, forums, video, reference sites, competitor comparison pages. You cannot prioritize off-site work without knowing which surface your category's engines actually draw from, and the Similarweb variance shows that differs enormously by industry.
Audit consistency, not just presence. The entity-footprint framing matters here. Three third-party sources describing your company three different ways is worse than one describing it clearly. Fix contradictions in how you are categorized, what you are said to do, and who you are said to serve before chasing new mentions.
Next two quarters
Treat review platforms and comparison sites as primary channels. If comparison-stage citations point overwhelmingly at third-party sites, then your presence, accuracy and volume on those specific sites is a direct visibility lever — and one most brands manage reactively, if at all.
Build the creator program around structure, not reach. The <1,000-view citation finding says the selection criterion is topical specificity and retrievability. That argues for many small, precise, well-structured collaborations over few large ones.
Start the coverage-volume work now, because it lags. If external mention volume is the threshold variable, there is no fast version. Digital PR, partnerships, integrations, conference presence, customer stories published on other people's properties — all slow, all compounding.
Ongoing
Re-measure across engines, repeatedly. Cross-engine agreement on brand recommendations is weak, and single measurements are noisy. The Perplexity social-share collapse from 31% to 13% in one quarter is the argument for continuous monitoring rather than a quarterly check.
How to measure something you don't own
The standard objection to off-site work is that you cannot measure it, which is why budget keeps flowing to the things you can. That objection is weaker than it used to be, and it is worth dismantling because it is the real reason this work does not get funded.
Four things are measurable today without access to anyone else's analytics.
Citation source mix. For a fixed prompt set, record not just whether you appear but what the engine cited to produce the answer — by domain and by source type. Tracked over time, this tells you which surfaces your category's engines actually draw from, which is the input to every prioritization decision below it. It also catches an engine changing its mix, the way Perplexity's social share moved from 31% to 13%.
Share of voice against named competitors, on category prompts. Not brand prompts. The 96%/89% split means brand prompts will report success regardless. The measurement that matters is: when a buyer asks the category question, who gets named, how often, and where do you sit.
Consistency of description across sources. Query several engines about what your company does, who it serves and what category it belongs to. Record the variance. High variance is a concrete, fixable finding that points directly at which third-party sources need correcting — and it is invisible to any page-level audit.
Mention volume and referring-domain growth over time. This is the slow variable, and the one the threshold research points at. It will not move in a month, which is exactly why it needs a baseline now rather than when someone asks why AI visibility has not improved.
None of that requires the cooperation of the sites doing the citing. All of it can be tracked continuously. The reason most teams do not is not that it is impossible — it is that it was nobody's job when the discipline was defined as a website exercise.
A note for B2B specifically
The pattern intensifies in B2B, for a structural reason worth naming.
Consumer categories have dense, high-volume third-party coverage: review sites, forums, creators, press. A consumer brand that has been around for a few years clears a mention-volume threshold almost as a byproduct of existing.
B2B categories are thinner. A mid-market software company with real revenue and satisfied customers may have a few hundred pages on the open web that mention it by name — analyst mentions, a handful of review-platform profiles, some conference listings, a few customer stories. Against a threshold in the low thousands, that is not close.
This is the mechanism behind a pattern many B2B marketers have observed and struggled to explain: the company is well known to its actual buyers, wins deals when it is in the room, and simply does not appear when an assistant is asked to list options. The buyers know you. The corpus does not.
The remedy is not more product pages. It is deliberately manufacturing the independent coverage that a consumer brand accumulates passively — analyst relations, review-platform programs, integration partner directories, co-authored content on other people's properties, customer stories published where the customer lives rather than in your own resources section.
Slow, unglamorous, and on current evidence the thing that determines whether you exist in the shortlist at all.
The caveats, stated properly
These are vendor figures. Tinuiti's numbers were published by a Tinuiti executive, in a piece advocating for a service Tinuiti sells. Victorious sells SEO and AI-visibility services, and its conclusion — that you need more third-party coverage — is one it is paid to act on. Neither has been independently audited.
The samples are small relative to the confidence of the percentages. 175 brands across five unnamed verticals does not support four-significant-figure claims like 99.99%.
Correlations near 0.45 to 0.49 are moderate. They are consistent with external coverage being one important factor among several, not with it being the cause.
The Tinuiti Q1 figures are already a quarter old, and the company has since released Q2 data. In a field where an engine's source mix can halve in ninety days, that matters.
Prompt phrasing alone can flip which brands appear in a category answer, which makes any single-run measurement fragile — including the ones these studies are built on.
What survives all of that is not a number. It is a direction, and the direction is corroborated by four independent groups using four different methods: the decisive surface for AI visibility is mostly not your website.
What to watch
Whether anyone replicates the 2,000-page threshold on a larger, independent sample. It is the most actionable claim in this batch and the one most in need of confirmation.
Whether engines keep changing their source mix quarterly. If Perplexity's 31%-to-13% social swing turns out to be normal rather than exceptional, then off-site channel strategy has to be portfolio-based rather than optimized for any one engine.
Whether owned content regains ground. Nothing here says owned content is irrelevant — it says it is a small share of citations today. Engine strategies on first-party sources could shift, and the brands that maintained both will be better placed than the ones that abandoned either.
The takeaway
Your website is table stakes, and you have probably already paid for it. If you have run a competent technical SEO program, most of the AI-readiness checklist is done, and paying again for it under a new name is a budget transfer rather than a strategy.
The part that decides whether you appear in a buyer's shortlist is happening on properties you do not own, cannot edit, and mostly are not measuring. It is slower work, it sits in a different budget, and it is the work.
See How AI Describes Your Brand
Discover where you stand in ChatGPT, Gemini, and Perplexity — and how Genmark AI can help you get cited.
Check Your AI Presence