We captured 12,264 of the searches AI systems write for themselves, across two markets and three engines. In many of them the machine investigated a brand by name, or described it accurately, and still recommended somebody else.
If your brand does not appear when a buyer asks ChatGPT to recommend a supplier, you probably assume one of two things. Either the machine has never heard of you, or your marketing is not where it needs to be.
We captured 12,264 of the searches AI systems write for themselves, across two markets and three engines: 177 category questions, each run five times, on ChatGPT, Claude and Gemini. In many of them the machine investigated a brand by name, or described it accurately when asked, and still recommended somebody else.
How we see those searches: ChatGPT, Claude and Gemini each return, alongside the answer, the list of web searches they ran to build it. We keep that list on every run and count the company names in it. This dataset was captured on 23 and 24 July 2026 on GPT-5.4, Claude Sonnet 4.6 and Gemini 2.5 Pro, the current versions at the time. Perplexity reports only how many searches it ran, and Google's AI Overviews and AI Mode report none, so those surfaces are outside these numbers.
That gap, between being investigated and being chosen, does not show up in any visibility score.
In the recommendation questions we tested, the engine rarely ran the user's words as a search. It wrote its own searches, usually several, and ran those instead.
Those self-written searches are the interesting part, because they contain brand names nobody typed. On 52% of the 89 category questions we tested in one market, and 66% of the 88 in the other, the engine's own searches named companies the user had never mentioned. The searches suggest the engine can arrive at retrieval with candidate companies already named.
All three engines do it, at different rates. Claude's own searches named companies most often in both markets, Gemini next, ChatGPT least.
Being named in those searches usually leads to a mention. Across everything we captured there were 438 cases where the machine's own searches named a competitor on a question. In 87% of them that competitor was mentioned in the answer. The machine usually mentions what it investigates. The other 13%, 56 cases, were named in the machine's own searches and then left out of the answer. Those 56 are the rejection cases we come to below.
You can have an excellent page and still never be one of the candidates competing to be in the answer.
One company in our dataset shows this more clearly than any theory could. It is the same company, the same website, the same people, selling the same service either side of a border.
In its home market it is the machine's first thought. It appears in the engine's own searches on 45 of 88 category questions, 222 times across repeated runs. Claude did most of that naming, 159 of the 222, Gemini 55, ChatGPT 8. In the neighbouring market it enters the machine's own searches once, across 89 questions and five runs of each.
Being named mattered. On the 45 home-market questions where the company entered the machine's own searches, it appeared in the answer every time. On the 43 where it did not, it appeared in 30.
The engine knows the firm in both places. Ask about it by name in the weaker market and you get an accurate, complimentary description. It almost never surfaces in the machine's own research when the category question is asked.
The obvious objection is that every firm is better known at home, and a company that has not launched in a market cannot expect to be recommended there. This company sells the same service in both markets from the same website, and when asked about it by name the machine answered every brand question correctly in both. Knowledge is the same on both sides of the border. Association is not. At home the machine connects the firm to the category and goes looking for it. Across the border it knows the firm and does not make the connection. A visibility score would report the weaker market as unknown. The searches say something more useful, and more fixable: known, described correctly, not considered.
A second company, in a completely different industry and a third market, showed the same home-market pattern: present in the engine's own searches on 36 of 65 questions, 235 times across repeated runs, with the engines the other way round: ChatGPT 180 of the 235, Gemini 31, Claude 24. For that company we have no second market to compare, so the border contrast rests on the first firm alone.
The machine's picture of the company is fine in both markets. The answer turns on whether the company comes to mind when the buying question is asked.
We then took every case in our data where an engine's own searches introduced a brand, investigation visibly followed, and the brand was absent from the final answer. There were 56 such brand-question pairs. That is 13% of the 438 cases where the machine's own searches named a competitor on a question. The 56 come from both markets and the second company's dataset.
Of those 56, we went back through five in full, following the machine's own searches, the pages it cited and the answer it wrote, to see where each brand fell out. Three patterns came up. Five is a small number, and we have not yet sorted the other 51 by pattern, so read these as early patterns rather than settled findings.
One of those searches, verbatim, with the firm's domain masked: site:[firm].co.uk virtual ciso. That is ChatGPT going into a specific firm's own website to check for evidence of a specific service before deciding whether to name it.
Investigated directly, and nothing was carried into the answer. One of the world's best-known professional services firms was probed directly: the engine wrote a search into that firm's own website, looking for evidence about a specific service. The firm's site then appears nowhere in the citations, and the firm is absent from the answer. The engines built their answers from smaller specialists whose pages state that service plainly.
Compared with specialists, and dropped from the shortlist. A large, listed company was swept into comparison searches alongside genuine specialists, on a question that asked for specialists. The comparison searches ran, the final answer omitted the brand, and the answer itself states the filter it applied: smaller specialist firms rather than large integrators.
Read and cited, and still not chosen. Worry about this one. On a question asking for a very specific kind of provider, a well-known company was investigated by name, and its own pages appear in the answer's list of sources. The brand is named in no answer. The competitor that wins the question offers something built precisely for what was asked; the investigated brand's pages described something more general.
The machine used the pages and then named a rival.
Most AI visibility measurement answers one question. Does the brand appear, yes or no. Appearing is the last step of a much longer sequence, so a no tells you nothing about which step failed. Some tools now show the machine's own searches as well. Seeing the searches does not tell you where one brand dropped out on the way to the answer.
Here is the sequence we can now watch. We call it the Decision Trace.
The Decision Trace
How an AI recommendation is decided
SearchIntel reads the decision that produced it.
We observe stages one, two and four directly. Stage three, evidence, we read from what the answer cited and from how it explains its own choice, so it is an inference and we label it as one.
SearchIntel · The Decision Trace
Failing at candidate formation and failing at selection look identical in a score, and they need opposite responses. No amount of on-page work will fix a brand the machine never nominates, and the weaker market has exactly that problem. The company whose pages were cited and passed over has a different problem: the pages were found and read, and the answer chose a rival whose pages matched the question more exactly. The first is a positioning problem, the second an evidence problem, and the same zero in a visibility score covers both.
None of the fixes are new. Positioning, authority, evidence, content that answers the buyer's actual need: good search strategists have argued for all of it for years. Until now you had to infer which one a brand needed from the outcome. A Decision Trace reads it from the decision itself.
How to read your own zero.
Absent from the machine's searches on category questions, but described accurately when asked by name: the machine knows you and does not think of you for the need. Look off your site first, at the directories, reviews, comparisons and third-party pages that put your name next to the category.
Named in the searches, then absent from the answer: the machine looked and did not carry what it found. Check whether your pages state the specific service or need, in the words the search used.
Cited in the sources, still not named: your pages describe something more general than the question asked. The rival that won matched the need exactly.
Your brand can be well known, accurately described, actively investigated by the machine, cited in the answer's own sources, and still left out of the recommendation.
Each of those is a different failure, and a visibility score of zero looks the same across all of them.
So the question worth asking is not "do we appear?". It is "where are we losing the decision?". We built the AI Decision Review to answer it: take the buyer questions that matter commercially, establish where in the sequence the brand drops out, and match the intervention to the failure instead of guessing.
Treat the score as the first question you ask, never the last.
We can observe investigation and selection. We cannot inspect the model's reasoning, so every explanation we offer is an inference from how the machine worded its searches, what it cited, and how the answer explains its own choice. We label them that way.
Two limits work against our own argument. Our candidate-formation measure undercounts: in one market, 30 of the 43 questions where the brand made the answer showed no brand-named search at all. The brand was selected without any brand-specific search we could see. And answers move on their own. We ran the second company's 65 questions on 4 June and again on 12 August 2026, having changed nothing ourselves, and six changed anyway, roughly 9%. Between those two dates competitors published, indexes updated, and all three models moved a generation: GPT-5.4 to 5.6, Claude Sonnet 4.6 to 5, Gemini 2.5 Pro to 3.5 Flash. We treat that as background movement any improvement claim has to beat, ours included. Our September test is pre-declared for that reason: treated and untreated questions, with pass and fail conditions written down before anything ships. We will report what it finds, including if it finds nothing.
Google's AI Overviews and AI Mode are in our scans but not in these numbers, because they expose none of their searches. Whether candidate formation works the same way there, we cannot yet see.