The AI
Visibility
Report

Every category we measure behaves the same way. A buyer describes a problem, an engine returns three to five names, and the comparison set is decided before anyone contacts anyone. If your name is not in that set, the loss never appears in a report. There is no bounce, no lost session, no line in the funnel. Demand simply goes elsewhere.

Two numbers frame the size of it. In over a quarter of our 204 measurements the brand was named in fewer than one answer in ten (Figure 4). And on the questions buyers ask before they know who you are, brands are named 28.4% of the time versus 77.3% when the question already contains their name (Figure 5). Most reported AI visibility is the second figure, and adding branded questions to a set lifts it without anything changing about the brand.

The uncomfortable part for planning: this is not a website project. The sources engines cite most are communities, review platforms and reference pages, not brand sites (Figure 7). And positions move when models update, without notice (Figure 10), so this is a standing operating cost rather than a one-off fix. The question for the board is not whether to believe in AI search. It is whether you can currently answer, with evidence, how often your company is named when your buyers ask.

Some of the questions your buyers put to an AI already have your name in them. Most of them do not. The first group tells you whether you are recognised. The second decides whether you get found at all. Across our corpus the two are nowhere near each other: brands were named in 77.3% of answers to questions carrying their name, and 28.4% of answers to questions that did not (Figure 5).

Reported as one number, those two collapse into a figure that lands wherever the question mix puts it. That is what makes a single AI visibility score unsafe in a board deck: it moves when the questions move, and nobody in the room can see which of the two it is describing.

Then the engine problem. Gemini names brands in 41.9% of answers, Claude in 31.0% (Figure 1), and when a brand is visible anywhere the engines disagree 56.6% of the time (Figure 3). A vendor quoting one number without naming the engine has chosen an engine, and with it a number.

So where do the answers come from? The most-cited domains that clear our disclosure rule are communities, social platforms and review sites (Figure 7), and answers that failed to name the brand actually cited more sources on average, not fewer (Figure 9). Brand-identifying domains are excluded from that table, so read it as a view of the third-party layer. What the citation counts do not show is why an engine chose someone else.

By the time a conversation starts, an engine has usually already returned three to five names with reasons attached. The comparison set was built without you, and the buyer treats it as neutral because it did not feel like advertising.

This explains something field teams notice constantly: two prospects in the same week describe the category completely differently, and neither is misremembering. When a brand is visible at all, the engines disagree with each other 56.6% of the time (Figure 3). Different buyer, different engine, different shortlist.

Two things to use directly. When a prospect says they checked and you showed up, ask which engine - the odds are better than even that another one disagrees. And when they say AI got a fact wrong, that is a traceable source problem, not an opinion, which turns an awkward moment into a credible next step (Figure 7).

Most brands in this data are trying to get into answers. Platforms with deep inventory and category pages have a different starting position: engines already read them to answer questions about an entire category. Listing and dealer marketplaces sit above the corpus average at 36.5% visibility across 19 brands (Figure 12), and the pattern behind that is structural, not promotional.

That changes the product question. It is not only 'are we visible', it is 'is our content liftable'. Engines quote what they can parse cleanly: a question answered directly, facts in stable structure, pages that do not require a session to read. Anything that hides inventory behind interaction is invisible to the layer that now recommends it.

Two operational notes. Improvements land unevenly across engines because each reads a different source mix (Figure 1), so measure all four before and after a release. And model updates can move a number without any change on your side (Figure 10), which is why release measurement needs a frozen question set to be worth anything.

Retrieval, not ranking. Engines assemble an answer from sources they fetch and trust, then return three to five names. Technical health, useful content and third-party references all still feed that process, so the fundamentals carry over. What does not carry over is position thinking. There is no page two to climb from.

Treat the four engines as four retrieval systems, not one surface. The spread between the most and least generous is 10.9 points (Figure 1), and verdicts split more than half the time (Figure 3). Optimising against whichever engine you personally check is the fastest route to being confidently wrong.

The most useful diagnostic in this report is the citation layer. Answers where the brand was absent carried 10.26 cited domains on average against 9.07 where it was present (Figure 9). Absence is not an indexing problem. Pull the sources behind the specific answers you lose, and that list is your work order - mostly off your own domain (Figure 7).

For the first time there is a measurable answer to a question comms has always argued about: which coverage actually influences what buyers are told. Across 22,111 answers carrying citations, the sources cited most often are communities, social platforms, review sites and reference pages (Figure 7). Brand-owned material sits well below them.

Tone is not the risk here. Across classified mentions the unfavorable share is small on every engine (Figure 6). The risk is absence, and absence does not trigger any alert in a monitoring tool. A brand can hold a spotless sentiment score and be missing from the answer entirely.

The most actionable pattern: engines can hold different versions of a story at the same time, because they read different sources. Correcting a fact in one place does not correct it everywhere (Figure 3). Recurring errors almost always trace to a specific citable page - a stale profile, an old article, an unmaintained directory entry.