Ninety-four percent of the brands AI cited while people were researching a category were also named when people asked which company to choose. For brands AI never cited at the research stage, the number was 50 percent. That is the headline finding of Victorious’s Q3 2026 Search Report, and it is the cleanest outside measurement yet of something operators have suspected for a year: the recommendation is decided earlier than the recommendation question.
Three tiers, and they are not the same job
AI visibility comes in three tiers. At the bottom, a brand is mentioned — its name exists somewhere in the answer’s world. In the middle, a brand is cited — its own content is used as a source while someone is learning how a category works. At the top, a brand is recommended — named when the buyer asks who to choose. Most reporting lumps the three together. The Victorious data says they behave like a ladder: the middle rung predicts the top one.
A citation is not a compliment. It is the platform treating your page as evidence. When a model answers “what should I check before choosing an insurance provider?” and builds that answer on your content, it has filed your company under sources I trust on this subject. The study suggests that filing decision carries forward into the next conversation, where the question is no longer “how does this work” but “who should I hire.”
What was actually measured
Victorious followed 161 brands across five industries — ecommerce and retail, financial services and insurance, healthcare, legal services, and SaaS — and collected about 16,500 AI answers at the end of September 2026. For the research-to-recommendation test, it used Ahrefs Brand Radar data: real questions people in the US asked AI platforms, sorted into research questions (“how something works or what to look for”) and consideration questions (“who to choose”). Six platforms that search the web supplied the citation data.
Read the finding the way Victorious itself reads it: as a correlation, not a cause. Bigger brands get cited more and named more, so the researchers checked whether size alone explained the gap. It explained part of it. After accounting for how often AI mentions each brand overall, cited brands were still more likely to be named. Victorious is now running a weekly panel through December to test whether research citations show up before the recommendations they predict. Until that lands, the honest statement is: citation and recommendation travel together, and the citation is the one you can work on directly.
The second number is the one to plan around
On July 1, forty brands in the study had never been named on a single consideration question. By October, 81 percent of the ones already cited on research questions had broken through and been named at least once. In the group with no research citation, 21 percent had. The groups were small — 21 brands against 19 — so treat the exact percentages as a direction, not a promise. The direction is not subtle: brands entered the recommendation tier through the research tier, at roughly four times the rate of brands that tried to arrive unnamed and uncited.
Meanwhile the room got smaller. AI named 42 of the study’s brands in the second quarter and 29 in the third — 15 dropped out, 2 came in. The amount of third-party web presence a brand needs for an even chance of being named rose from about 20,000 mentions to somewhere between 24,000 and 28,000. Visibility at the top tier is not a possession. It is a lease, and the rent went up.
The research shelf itself is being rearranged
Brands are also losing ground inside the research answers. Across the twenty research questions Victorious tracks, brand sites’ share of citations fell from 65.2 percent in Q2 to 55.5 percent in Q3, while government sites climbed from 4.2 to 11.7 percent and association sites from 3.9 to 9.8 percent. And in every industry, the ten most-cited websites took a bigger share of all citations — in legal services, about half. Your research content no longer competes mainly with other companies’ research content. It competes with the regulator, the trade association, and a short list of incumbents the platforms already trust.
That rearrangement is why “publish more explainers” is not a strategy. An explainer that reads like a brochure loses the citation to the association page that reads like a reference. The research answer that gets cited is the one a careful professional could rely on: specific numbers, current dates, named sources, no sales language. Write the page you would want quoted back to you in someone else’s answer.
What we do differently knowing this
- Track the questions before the question. For every “who should I choose” query you care about, name the two or three research questions a buyer asks first, and measure whether your content is cited in those answers — not only whether your name appears in the buying answer.
- Report the tiers separately. Mentioned, cited, and recommended are three different numbers with three different fixes. A dashboard that blends them hides the one that predicts the others.
- Earn placement inside trusted sources, not just on your own site. Association guides, trade publications, and directories are taking citation share. A contribution there puts your name inside a source the platform already relies on.
- Measure on a rhythm, because citations drift. Victorious found that asking one platform the same question twice reuses only 25 to 74 percent of its sources, and three months apart the overlap falls to 6 to 16 percent. A single screenshot of an AI answer is an anecdote. A monthly reading is a measurement.
- Defend the research shelf you already hold. Fifteen brands fell out of the named set in one quarter. If AI cites your explainer today, that page is an asset with a maintenance schedule, not a finished job.
Where Tygart Media sits in this
We run our own fleet on the same logic this study measures. Tygart Media is an AI-first agency and publishing house working out of Tacoma and Puyallup, and our sites are measured on Bing and Copilot citations the way other shops measure rankings. Our October map of tygartmedia.com counted 3,229 published posts, and our most-cited page — a pricing explainer — carried about 96,000 citations in a September Bing export. That is research-tier weight, built on purpose: explainers written to be used as evidence. The Victorious numbers say that weight is exactly what the recommendation tier is drawn from. We wrote earlier this year that cited is not chosen. This study sharpens the point: cited is not chosen — but it is the qualifying round, and almost nobody gets chosen without running it.
The practical takeaway fits in one line. If you want an AI platform to say your name when a buyer asks who to choose, first give it a reason to cite your work when that buyer was still asking how the job works. The explanation is the audition. The recommendation is the callback.
Get The Signal
What we actually ran, what it measured, and what it means for a restoration shop — every week, including the flops.
Leave a Reply