What Keeps a B2B Company in AI Answers: The VizzEx AI Citation Study, in a Nutshell

VizzEx Longitudinal B2B AI Citation Study · Final report

Your buyers are asking AI engines the questions they used to bring to your sales team. By the time they reach out, many of them have already built a shortlist from those answers. So the question for a marketing leader isn’t whether your company showed up in an AI answer once. It’s whether you’re in the answer week after week, and what you can do about it when you’re not.

That’s what the VizzEx Longitudinal B2B AI Citation Study measured. For 12 weeks, July 25 to October 28, 2026, we tracked 10 company websites: 8 that used or were observed under the VizzEx approach, and 2 famous marketing publishers as untreated controls. Each site had its own set of questions, mostly the specific questions its buyers ask, 1,431 in all. Every week, we checked those questions on Google AI Overviews, ChatGPT, Claude and Perplexity, and stored all 59,645 answers.

The full report is 50 pages. This is the study in a nutshell.

Download the full report (PDF)

How this study is different

There are hundreds of “which sites does AI cite” studies. Most take one snapshot across thousands of sites, consumer, local and B2B mixed together, and report what the cited pages have in common. That shows a pattern across the web. It doesn’t tell you why your company was cited, why it stayed cited, or why it disappeared.

This study was built to answer those questions:

  • The same questions, every week. 12 consecutive weekly checks, not one snapshot or a monthly average.
  • Traced to the post. Each of a company’s citations was joined to its own post and that post’s features: headings, semantic relationship links, structural symmetry, FAQ sections, Google rank, and whether other engines cited it too.
  • The losses, not just the wins. When a company wasn’t cited, we looked at the pages the engine cited instead and compared them with the company’s own post, on the same question, in the same week.
  • Controls. Two famous marketing publishers, untreated, measured the same way.

1. AI answers change every week

From one week to the next, 15–28% of cited questions changed, depending on the engine. Google AI Overviews changed the most. A first citation was still there the next week only about half the time on Google, Claude and Perplexity, and 17% of the time on ChatGPT. If you want to understand which citations are most likely to vanish and why, the patterns behind that drop-off are not random.

That’s why a one-time check of your AI visibility tells you so little. A citation you can’t count on is hard to plan, budget or report against. The return comes from citations that hold, and that come back when they drop. (Report, p. 12)

2. Some citations hold, and you can see why

Against that weekly change, 39–58% of the citations the companies won held for 4 or more weeks in a row, depending on the engine: Claude 58%, ChatGPT 54%, Perplexity 48%, Google AI Overviews 39%. On every engine, between 13 and 84 questions stayed cited at every check for 10+ weeks.

We joined every won position, 1,499 question-and-engine pairs, to the factors around it. These are the ones that predicted holding:

  • Several connected posts could answer the question: held 35 to 57 points more often (median +42), better on all 8 study sites. On one site, 85% of those questions held, against 28% of questions answered by a single post.
  • Other engines also cited the company: +33.5 points (median).
  • An FAQ section on the post: +17 and +21 points, in the two sites with clean comparisons.
  • The company was named in the answer: +13.5 points (median). (But the company was named in the answer is not the same as the company being recommended: the question the buyer asks determines which of those happens).
  • More semantic relationship links to other posts in the blog: about +10 points (median), positive on 5 of 6 sites, and 27 to 39 points on longer posts.
  • Higher-volume questions were harder to hold: −17 points. They’re the most contested.

Structure works as a threshold. Posts on the study sites outscored the page they were up against on structural symmetry in 89% of 1,163 head-to-head comparisons, but once a post was past the bar, better structure added only about 5 points. Heading density works as a range: about one heading for every 100 words. (Report, pp. 16–19)

3. Information only you can publish gets you in. Volume and fame didn’t.

Coined terms, named frameworks and named client stories earned citations fastest, on every site. This is what VizzEx calls the Unique Information Delta: information a buyer can’t get anywhere else.

Volume didn’t do it. One site published 128 new posts that weren’t connected to the rest of its blog. They drew 1 of its 179 citations at the end of the study. On another site, 11 new posts connected to related posts drew 27 citations, and 9 of the 11 were cited.

Fame didn’t do it either. A famous marketing publisher with 96,060 other domains linking to it got no ChatGPT citations in 12 weekly checks. On that publisher’s own keywords, the engines cited 125 LinkedIn posts by working practitioners and none by its famous founder. (Report, pp. 13–16, 36). We wrote up that control-group result in detail, because it surprised even us: a publisher with 96,060 domains linking to it still got no ChatGPT citations on its own keywords.

4. The company keeps the citation when a post rotates out

Engines don’t only drop a company; they rotate among its posts. On questions several of a company’s posts could answer, when an engine stopped citing one of them, it moved to another post from the same company 79–94% of the time. Across 1,296 of those switches, the new post was linked to the old one by a semantic relationship link 63% of the time, against 33% if the engine had picked at random.

This is the VizzEx GEOMesh™ at work: posts connected by semantic relationship links, so when one rotates out, another of yours takes its place. The famous publishers have no semantic relationship links between their posts. When one of their posts dropped out, they lost the citation for that question 79 times in 90. (Report, p. 21)

5. Lost citations come back, mostly without edits

A dropped citation is usually not a lost one. On the sites using VizzEx, lost citations came back 73% of the time (134 of 184). The famous publishers’ came back 41% of the time (9 of 22).

Most came back without anyone touching the post. Posts edited after the loss came back 75% of the time; posts left unchanged, 73%. Google’s own updates were part of the picture: of the 99 losses on Google AI Overviews, 56 came during an announced spam update or the late-summer core update. (Report, pp. 21, 33–34, 47)

6. Four engines, four buyers

The engines aren’t one market:

  • Google’s top 10: 43% of the pages Google AI Overviews cited ranked in Google’s top 10 for the question. On Claude, 20%. On Perplexity, 16%. On ChatGPT, about 6%. Ranking well in Google isn’t a plan for AI visibility. We looked at what that gap means for how you build visibility, in authority isn’t the price of admission for AI search.
  • Social platforms: Google AI Overviews cited a social platform in 71% of answers; ChatGPT in 1%.
  • Who’s first: Google AI Overviews and Perplexity were usually the first to cite a company on a new question; ChatGPT was the slowest. (Report, p. 24)

7. When you’re not cited, who is?

On Google AI Overviews at the final check, when a company wasn’t cited on its own buyer question, other companies’ and organizations’ sites took 71% of the citations. YouTube took 9%, LinkedIn 5%, other social platforms 4%, and Reddit 3%. The competition for your buyer’s question is mostly other companies’ content, not Reddit threads. (Report, p. 10)

What to do with this

The report ends with five things a B2B company can do now, each tied to what the study measured:

  1. Start from your buyers’ specific questions.  (If you want the thinking behind why that first step matters most, start from your buyers’ specific questions walks through why chasing high-volume head terms is the wrong place to begin).
  2. Publish what only you can.
  3. Put your name where each idea is taught.
  4. Connect your posts.
  5. Measure holding, not snapshots.

The full report explains each one and what the data says about it, and adds a checklist for getting ahead of your competition (Section 9, p. 27).

Get the full report

This post covers the headlines. The 50-page report has the rest: the factor-by-factor evidence, the losses compared page to page, the market’s beliefs tested against the data, the controls in detail, how Google’s own records line up with what we measured, and the method behind every number.

Download the full report (PDF)

Kim Albee is co-founder of VizzEx with Carolyn Holzman. The VizzEx Longitudinal B2B AI Citation Study tracked 10 company websites on Google AI Overviews, ChatGPT, Claude and Perplexity every week for 12 weeks.

Written by: — Founder

Founder of Genoo (B2B marketing automation) and co-founder of CampaignCoach.ai and VizzEx.ai — two commercial AI products built from 4.5 years of hands-on AI system development.