AI visibility: what should an agency actually fix first?

How AI Answers Questions About Brands · Part Four of Four

Part Four: The Agency Playbook

630 pages that never name the brand · 19 pages all five engines trust · 2,000 questions asked twice

First published September 2, 2026 · updated September 3 with a fresh full run — the study now measures five engines, adding Google’s AI Overview.

The series
Part One · The Retrieval Study — do the engines even search before answering?
Part Two · The Source Study — what the engines actually read when they do
Part Three · The Reddit Pipeline — the most negative layer of the study, and the two engines that carry it
Part Four · The Agency Playbook — you’re reading it

It starts with a screenshot from the client. An AI answer that recommends a rival, or repeats a complaint, or simply never mentions them. Then the question every agency has learned to brace for: “What do we do about this?”

Data isn’t the problem. Good tools expose real GEO and AEO numbers now — Semrush and Ahrefs among them — and anyone can see who’s mentioned where. What’s been missing is the part the client is actually asking for: the order of operations. What to fix first. What each fix is worth. And what number will move next month to prove it worked.

Parts One through Three of this study measured the terrain: which engines search, what they read, and what Reddit does inside their answers. This part is about what those measurements mean for that order of operations — what kind of work the data actually supports, and what to measure after it.

Start with the split most plans miss. Across the 1,500 answers from the three engines that report their own searching, answers where the engine searched the web got cited pages behind them — and a citation went to the brand 15.3% of the time. Answers where it didn’t search? Zero citations. Not few — zero, in all 395 answers. Searching is the only road to a citation.

But here’s the twist: those no-search answers still mentioned the brand just as often — 68.9%, against 68.1% when the engine searched. Retrieval buys links, not love. Being named lives in a different place than being linked — and a plan that funds “AI visibility” as one line item is quietly buying only one of the two.

Part Two audited 2,357 cited pages. Sort them by how many engines lean on them and a species split appears. Pages cited by three or more of the five engines — 26.9% of the set, 633 pages, only 19 of them carried by all five — are overwhelmingly ranked lists, seldom negative (9.5%), and usually name the brand. Single-engine pages are the opposite animal: forum- and video-heavy, 20.0% negative, and brand-absent nearly half the time.

That split is an ordering principle. The consensus pages are the shared load-bearing wall of every engine’s answer — fix those first. For YETI, one page — a cooler lab test on outdoorgearlab.com — was cited 88 times this wave, by four of the five engines, Google’s Overview included. It does more work in YETI’s answers than dozens of pages combined. Peak Design’s own warranty page is a four-engine consensus source — the Overview among them — proof that an owned page can make the list. And its Trustpilot page is cited by all five engines and open to none of them’s auditors: a person in a browser sees it fine, but every automated reader hits a wall there, ours included.

Here is the study’s biggest opportunity number. Of the 1,884 cited pages the instrument read, 630 — 33.4% — never mention the brand whose buyers they’re answering. These pages already rank, already get cited, already sit inside your category’s answers. They simply don’t include you.

The spread across brands is the proof it’s fixable: Gymshark’s reading list is brand-absent on 48.6% of pages, AG1’s on 45.9%, OLIPOP’s on 45.8% — while Patagonia’s sits at 19.8%. Same engines, same kinds of pages. The difference is presence on pages someone earned. For an agency, each of those 630 pages has a known URL, a known author, and a known reason to update. That isn’t a statistic — it’s a call list.

When the reading layer recorded which competitors appear on cited pages, the sets came back category-locked. Only two names — Nike and Adidas — recur across multiple brands’ reading lists. Everyone else’s competitive field is a handful of names that never leave the category. The engines have already decided who your client actually competes with — and it’s a shorter list than the one the client worries about, and often a different one.

How often does an engine cite the brand’s own website in that brand’s answers? The rates are low everywhere, and they’re an engine property like everything else in this series:

• Perplexity29.4%of its 500 answers cite the brand’s own site
• Google AI Overview23.7%of the 367 questions where it appeared
• Gemini16.6%
• Claude13.6%
• ChatGPT3.6%and for one brand — Hoka — 0 of 50, in both waves

Pair that with Part Two’s finding that 53.8% of cited own-site pages wouldn’t even open for a machine reader, and the shape of the own-site job is clear: modest ceiling, mostly unclaimed. And on the demand side, the biggest hole in the study: the long-tail lane — “best X for my situation” questions — where brands get mentioned in just 36.8% of answers overall, and AG1 in 1.7%. The questions closest to a purchase are the questions brands are most absent from.

None of this matters if the target citation won’t hold still. So the study ends with the measurement everything else depends on. When the first wave ran, we re-asked every one of its 2,000 assistant questions, same engines, same wording, within a twenty-four-hour period — and compared everything. The answer mention rates barely moved: 69.0% became 68.3%. But underneath, the engines rewrote roughly half their own search queries — and what happened to the cited sources splits cleanly by engine:

Like searching, like Reddit, like everything else in this series: consistency is an engine property. This measures a single day — how the picture decays over weeks is the follow-up measurement, and it isn’t published until it’s run.

And this very update corroborates the experiment at a two-day interval: the fresh wave behind this article re-measured the whole study with a fifth engine added. ChatGPT’s search rate drifted from 31.4% to 33.2%, Perplexity halved how many sources it displays, a second Reddit carrier appeared — and the aggregate mention layer barely moved (68.8% on the four shared engines, against roughly 69% in wave one). The surfaces churn; the aggregates hold. Both halves of that sentence are the finding.

For an agency this number is the license to work. The reading list holds still enough to be worth fixing — and moves enough that every fix needs a re-measurement behind it. Where recurrence is high, a page you win stays won. Where it’s low, visibility is a monthly reading, not a trophy. The instrument this series ran on treats that as a feature: StyleForge’s AI Visibility Pulse versions every run, so a second pulse reports its variance against the previous one — every engine rate, share-of-voice position, and question type carrying its change:

Everything in this series ran on StyleForge’s AI Visibility Pulse, and this is the part where I say plainly what it is and why we built it this way. Plenty of tools now report GEO and AEO data, and some of it is excellent. Our premise is narrower: arm the agency with findings it can put in front of a client and act on together. The reports exist so agency and client can work from the same document toward the same goal — growing the brand’s visibility with AI assistants — with every finding pointing at a fix, and every fix pointing at the number that will show it worked.

In practice: the question pack is drafted for you by AI — or drag in your own list from a template — with the competitors, question types, and weighting under your control, so an audit can be an investigation, not a preset. The citation audit reads the sources channel by channel, at the depth you choose, down to full YouTube transcripts. Every pulse is versioned, and a second run reports its movement against the last one at the question-type level. And the client report — 200 to 250 pages, white-labeled to the agency, section-selectable — closes on a recommended-actions page where every action names an owner and the number that will move next pulse — the same discipline this series has practiced throughout: no claim without a number, no action without a re-measurement behind it.

The complete audit for every brand in this series is published below — every question, every answer, every cited page read. Click any brand to open its full report.

Each of these reports is white-labeled — generated under the agency’s own name and presentable straight to the client, and you control which sections every report includes. The agency name on these ten is a stand-in.

I’ve been building websites since 1995 — hundreds of them, hand-coded, optimized, fussed over — and I watched the whole SEO industry grow up around that work: the early tools, the guesswork, the slow arrival of real measurement. That experience is why I added an AI visibility tool to StyleForge: agencies deserve to see what’s actually happening under the hood of these answers, and to have a way to act on it with their clients — so no brand finds out secondhand how the assistants talk about them.

Just as the analytics tools of early SEO evolved and got better year over year, I think that’s exactly what’s happening with GEO right now. This series is my contribution to it: measurements on the table, the full record published.

If you’d like to see your own brand — or your clients — through the same lens, you’re warmly invited to try the AI Visibility Pulse, and whether you do or not, I genuinely welcome your feedback.

Glenn · Founder, styleforge.io · September 2026

Every check in the study, written out in full.

Related resource: the complete guide to earned media — definition, examples, and a working earned media value calculator. The coverage this series measures is earned media; that guide is how to value it.