Across about 26,400 AI answers we logged over four months, the brand the question was actually about showed up in roughly one in six of them. The other 84 percent named a competitor, named no one, or answered the question without the brand anywhere in sight.
That is the number most visibility conversations skip past. People ask whether ChatGPT “knows” their brand, run one prompt, see their name, and relax. One answer is not a measurement. When you watch the same brands across thousands of answers and five different assistants, the honest picture is that being left out is the default, and the engines disagree sharply about who they will cite.
How we measured it
We logged about 26,400 answers from five assistants, ChatGPT, Claude, Gemini, Perplexity, and Google AI Mode, to category and buyer questions across a sample of brands in competitive categories, from late March to mid-July 2026. For each answer we recorded one of four outcomes for the brand the question was about: Cited (named with a link or an inline source), Named (named in the text with no link), Considered (pulled into the answer’s retrieval pool but not surfaced to the reader), or Missing. Every rate below uses answers as the denominator.
One caveat worth stating up front: these are brands in competitive categories, asked on questions where you would expect them to come up, not a random crawl of the web. Read the levels as directional for that kind of brand, not as a law that holds for every query on the internet.
The gap: named or cited in 16 percent, missing from 84
Pooled across every answer, the brand was Cited in about 13 percent, Named without a link in about 3 percent, and only Considered in about 1 percent. Add the first two and you get Brand Coverage of roughly 16 percent. The remaining 84 percent are answers where the brand simply was not there.
The 3 percent Named tier matters more than its size suggests. Those are answers that talk about your brand by name and give the reader nowhere to click. You are in the conversation and getting none of the referral, which is a different problem to fix than being absent entirely.
The engines do not agree on who to cite
Splitting the same answers by assistant is where the “AI is one thing” assumption falls apart. Claude named or cited the brand in about one answer in five, the most of any engine. ChatGPT and Google AI Mode sat at the bottom, roughly one in seven.
| Assistant | Present (cited + named) | Cited | Named |
|---|---|---|---|
| Claude | 20.3% | 17.2% | 3.1% |
| Perplexity | 17.5% | 15.4% | 2.1% |
| Gemini | 16.6% | 13.5% | 3.1% |
| ChatGPT | 14.1% | 9.4% | 4.7% |
| Google AI Mode | 14.1% | 12.1% | 2.0% |
ChatGPT is the interesting one. It has the lowest Cited rate of the five, under 10 percent, but the highest Named rate. When ChatGPT does bring a brand up, it is more likely than the others to mention it in prose with no link attached. If you only track linked citations, you undercount how often ChatGPT is talking about you.
A win on one assistant tells you almost nothing about the next. A brand that has clawed its way into Claude answers can still be invisible in ChatGPT for the same question, and the fix is not the same on both. This is why a single blended visibility score, one number averaged across engines, hides the thing you can actually act on.
The typical brand is worse off than the average
Averages flatter. The pooled Brand Coverage of 16 percent is pulled up by a handful of brands that show up almost everywhere. Look at the middle of the distribution instead and the typical brand was present in about 13 percent of the answers about its own category, with a spread running from near zero to roughly 79 percent.
So the gap is not a story about one or two weak outliers dragging a healthy field down. It is the median. Most brands in a competitive category are absent from the large majority of AI answers about that category, and a small number of well-established names soak up the visibility.
What we did not measure
Two things, stated plainly, because leaving them out would overclaim.
We cannot see the exact words a person typed to get an answer. We measure the answers, not the prompts behind them, so we can tell you how often a brand appears, not what a specific customer asked to make it appear.
We also do not report a “the model searched the web” versus “the model answered from memory” split, even though it is the obvious next question. Our current data cannot separate those two cleanly enough to publish a number we would stand behind, so we are not going to invent one. The mechanism below explains why the distinction matters even without a clean measurement of it.
Why the brand is missing, when your page ranks first
The uncomfortable part for anyone with strong classic SEO is that a page can rank first in Google and still never appear in an AI answer to the same question. Two documented mechanisms explain it.
First, the assistants do not always search. Whether to run a live web search is a per-request decision the model makes. OpenAI’s own developer docs say the model “can choose to search the web or not based on the content of the input prompt,” and Anthropic describes Claude using “its reasoning capabilities to determine whether the web search tool would help.” When an assistant answers from what it already learned during training, it never queries a live index, so your Google position is not an input to that answer at all. Your ranking cannot help on an answer that never looked.
Second, even when the assistant does retrieve, it does not just read the top result for your query. Google describes its AI Mode using a query fan-out technique that breaks one question into many sub-queries run in parallel, then synthesizes the answer from across those results. A page that ranks first for the head term can lose to pages that better answer the decomposed sub-questions. The academic work on retrieval and the knowledge boundary formalizes the first half of this: models lean on trained memory and reach for retrieval only when their internal confidence drops, because searching costs latency.
The stakes are not academic. Pew found that when a Google result carried an AI summary, users clicked a traditional link on just 8 percent of those pages versus 15 percent without one, and clicked a source cited inside the summary about 1 percent of the time. The answer is becoming the destination. Being the blue link under it is worth less every quarter.
What to do with this
Three practical reads come out of the numbers.
Measure each assistant on its own. A 16 percent blended number would have hidden that Claude cites you at 20 percent and ChatGPT at 14, which are different problems on different surfaces. Track Brand Coverage per engine, not as one score.
Go where the answer is built, not just your own page. Because so many answers are assembled from third-party sources rather than a live read of your site, the lever is usually the pages the engines already trust for your category, the review sites, roundups, and reference pages that keep getting cited. Being present there moves more than one more post on your own blog.
Treat a Named-without-link answer as its own fix. If an assistant is describing you in prose with no citation, the brand is known but not clickable, and the work is earning the linked source, not building awareness.
The wider point is the one this blog keeps coming back to: ranking in classic search and appearing in an AI answer are two different games, measured in two different places. We pulled apart how little the engines’ cited webs overlap in the engine-overlap study, and why an assistant reaches for a competitor instead of you in why AI recommends your competitors. If you take one thing from the 84 percent, it is that absence is the starting position, and you find out which engines you are missing by measuring them apart, not together.
