The State of AI Search, September 2026

Why this matters

When someone asks ChatGPT, Gemini, or Claude a buying question, such as who does accessibility testing or which company handles database consulting, the assistant answers with a short list of names. It does not show ten blue links. It does not show an ads auction. It picks, usually, a handful of companies and says them out loud. For a company that is not on that list, the effect is close to what happened to a business that stopped showing up on page one of search results a decade ago, except there is no page two to scroll to. There is just the answer the assistant gave, and the next question.

This matters to three groups differently. Marketing and growth teams need to know whether their category is one where AI assistants are already being consulted as buying advisors, because that changes where budget and content effort belong. Founders and operators need to know whether invisibility on one engine is a rounding error or a structural problem, because the fix is different for each. And anyone who has been sold a single "AI visibility score" needs to know whether that score could possibly be measuring something stable, given that the three engines are, as this study shows, not looking at the same thing.

The honest starting point is that nobody outside the model providers knows exactly how an assistant decides who to name. There is no public ranking algorithm to read, the way there once was a search algorithm to reverse-engineer. The only way to find out what these systems do is to ask them the same questions many companies would plausibly ask, at scale, and record what comes back. That is a measurement exercise, not a guess, and it has a specific shape: it tells you what happened during a defined window, for a defined cohort, under defined prompts. It does not tell you why, and it does not tell you what will happen next month if a model is updated. This study, covering 58,803 companies across 3 engines over 8,188 distinct prompts, is built to answer the narrower, checkable question: during this window, who got named, how often, and how consistently across engines. That narrower question turns out to have a clear, if uncomfortable, answer, and the rest of this page is about what that answer is and is not.

How the measurement works

The design question for a study like this is not "what do AI assistants think of my company", which is not answerable, but "what do AI assistants say when asked", which is. The method here is a prompt-and-capture design: write realistic buying questions, send them to each engine the same way a prospective customer would, and record whether a given company's name appears in the answer and, if so, in what position.

The run. Over a run window from 2026-09-04 to 2026-09-25 (2 boundary timestamps marking start and end), the study sent 8,188 distinct prompts across 3 engines, each running a single pinned model version (one version per engine, 1 tracked per engine), and collected 536,301 rows of raw output. Of those, 536,301 came back as usable answers, a completion rate that means the 536,301 figure used throughout this study is, for practical purposes, the whole attempt, not a filtered subset.

The prompts. Prompts were split into four types, reflecting different stages of a buying journey: informational ("what is X"), navigational ("where do I find X"), commercial-investigation ("best X for Y"), and transactional ("who should I hire for X"). Each type returned enough answers to stand on its own:

The four rates sit close together, within a few points of each other, which is itself a finding: whatever drives whether a company gets named, it is not strongly a matter of which stage of the buying journey the question represents.

What this cannot see. A pinned model version per engine means the study measured a snapshot, not a trend. Model providers update their systems without notice, and a re-run in a later month could show different numbers for reasons that have nothing to do with any company's own actions. The study also cannot see why a model chose to name one company over another, only that it did. Mechanism is inferred in later sections, not measured directly.

Which models answered. Each engine was queried through its vendor's developer API, not through the consumer ChatGPT, Claude or Gemini apps, which add account history, personalisation and their own defaults. Almost every answer in this window came from the vendor's small, fast model tier responding from the model's own knowledge, with web search switched off: OpenAI's GPT-4o mini, Anthropic's Claude Haiku and Google's Gemini Flash Lite. In mid September Rankxa moved to newer models with web search switched on. Less than one percent of this edition's questions were answered on the new setup, so this edition remains comparable with July and August. The next edition will be the first on the new setup, and it will say so.

The headline result, and its limits

Across the full cohort of 58,803 companies, 56,194 of them, 95.6% of the cohort, were named by at least one engine at least once during the run window. Only 0 companies, 0% of the cohort, never appeared on any engine for any prompt. Put plainly: almost every company in this cohort is visible to AI search somewhere, some of the time.

That is the headline, and it is easy to misread. It does not mean every company gets recommended often, or prominently, or on every engine, or for every question a customer might ask about its category. "Named at least once, somewhere, during a three-week window" is a low bar, by design, because it answers a narrower question than the one most companies actually care about: not "can I ever be named" but "am I reliably named when it counts". Those are different claims, and this study is built to separate them rather than collapse them into one score.

The gap between the two becomes visible once you look at engine-by-engine visibility rather than the any-engine figure. On ChatGPT, 44.6% of the cohort appeared somewhere. On Claude, 40.5%. On Gemini, 32.4%. Each of those is meaningfully lower than the 95.6% any-engine figure, which only makes sense if a large share of companies are being named on one engine and not the others. The any-engine number is a ceiling, built from different companies clearing the bar on different engines. No single engine comes close to it alone.

This is the first honest caveat of the study: a company that reads "95.6% of companies are visible to AI search" and concludes its own visibility is fine has likely misread the number. The relevant question for any one company is not the cohort average but its own per-engine record, and the cohort average actively obscures the per-engine variation that matters most for deciding what to do next. The next section looks at why that variation is shaped the way it is, rather than being a smooth gradient.

Is visibility gradual or binary

A natural assumption is that AI visibility behaves like organic search visibility: a long, smooth gradient, where companies trade places a little lower or a little higher depending on effort, content, and authority. The data here does not support that picture, at least not as the primary shape.

The clearest evidence is the near-total absence of companies that are simply invisible everywhere. 0 companies, 0% of the cohort, never appeared on any engine. If visibility were a gradient with a hard floor that some companies sit below, you would expect a meaningful tail of permanently invisible companies. There is effectively none. Combined with 95.6% of the cohort appearing somewhere, the pattern looks less like a ladder companies climb and more like a lottery companies mostly win at least once, if you pull the lever across three engines and 8,188 prompts.

But "appears at least once" is a different axis from "appears consistently," and on the consistency axis the data is closer to binary than graduated, just not in the direction of in-or-out. Look at the by-category breakdown: every single category in this study, from Accessibility Testing Services (2,147 companies) to E-commerce Solutions (3,600 companies) to small categories like Fitness Digital Marketing (7 companies), shows a visible rate of 100%. That uniformity across wildly different category sizes and subjects is unusual. It suggests that at the category level, the binding constraint is not whether a category gets covered by AI assistants at all, categories are essentially all covered, but which specific companies within a covered category get named on which specific engine, and how often.

So the honest shape is layered: binary at the "does this category exist to the models" level (yes, almost universally), binary-ish at the "does this company ever get named" level (yes, for almost everyone), and much more varied, close to a real gradient, at the "how often, how prominently, on which engines" level. A company deciding what to do should not ask "am I visible" since the answer is almost certainly yes somewhere. It should ask where on the frequency and cross-engine-consistency gradient it sits, which is a harder, more useful question answered in the following two sections.

What the engines are reading when they decide

This study measures outputs, who got named, not the retrieval and ranking process behind an answer, so any statement about mechanism here is inference from pattern, not a direct observation. With that stated plainly, a few patterns are informative.

Position is a separate measurement from presence. Across 197,893 ranked mentions, the overall average position a company appeared at was 4.94. Engines were close to each other on this: ChatGPT averaged 5.08, Claude 4.97, Gemini 4.72. The closeness here, less than a quarter of a position apart, contrasts sharply with how much the engines disagree on who to name at all, covered in the next section. That contrast is itself informative: once an engine decides to name a company, it places that company in a broadly similar slot to where other engines would. The disagreement is almost entirely about the guest list, not the seating chart.

Category coverage looks like a floor, not a differentiator. Every category measured shows full, or effectively full, visible rates, including small categories with single-digit company counts like Tourism Digital Marketing (8 companies) and large ones like Real Estate Services (4,139 companies). This is consistent with models having been trained on, or retrieving from, broad web coverage of business categories generally, rather than having sparse knowledge of any particular niche. It means a company cannot explain away poor visibility by claiming its category is too obscure for the models to know about. The category is known. The company's specific standing within it is the variable.

What the study cannot see here. Without retrieval logs or source citations captured per answer (this run did not capture a usable set of cited sources), it is not possible to say from this data alone whether engines are drawing on recent web content, older training data, structured business listings, or some blend, for any given answer. A company that improves its public web presence this month and sees a visibility change next month has a plausible causal story available to it, but this study cannot confirm that story against the alternative, that the change was simply a model update unrelated to the company's own actions. Distinguishing those two requires repeated measurement over time with the model version held constant, which is part of why this study pins model versions per engine rather than tracking a moving target.

Where the engines disagree, and why one score is misleading

If a company is choosing between a glance at cohort-level visibility and a per-engine breakdown, this section is the argument for the per-engine breakdown. The engines do not agree with each other about who to name.

The agreement table. Among all pairs of engine-and-company combinations where at least one engine named the company, here is how often one, two, or three engines agreed:

Out of 178,767 such pairs, 88.1% were named by exactly one engine and not the other two. Only 3% were named by all three. That is close to the opposite of consensus. If three independent human experts were each asked "who would you recommend for this job" and agreed with each other only this rarely, nobody would describe them as measuring the same thing, and would instead say they were drawing on different information or different judgment entirely.

Why this happens. The most likely explanation, consistent with what this study can see, is that each engine's training data, retrieval sources, and ranking tendencies differ enough that they are effectively running independent searches rather than consulting a shared index. This is a plausible account, not a confirmed one. An alternative explanation is that the underlying truth, which companies genuinely deserve mention, is itself ambiguous or close to tied for many categories, so small differences in phrasing or sampling tip different engines toward different, equally defensible, answers. The way to tell these apart would be to repeat the same prompts against the same engine many times and see whether an engine disagrees with itself as often as engines disagree with each other. This study did not run that repeated-prompt test, so it cannot settle the question.

The practical consequence. Any product or report that claims a single "AI visibility score" for a company is compressing three disagreeing measurements into one number. That compression hides the fact that a company could be strong on one engine and absent on another, which is a different, more actionable, problem than being uniformly weak. A single score answers a question nobody asked, and obscures the one most companies actually need answered: which engine is the problem.

What a company on the wrong side of this should do

Start by locating the actual problem, not the symptom. Given how the data behaves, "I am not visible in AI search" is rarely the accurate diagnosis for any single company in this cohort, since 0 companies fell into that bucket. The accurate diagnosis is almost always more specific: absent on one engine while present on others, present but ranked low in position, or present only for certain prompt types.

Step one: check per-engine, not cohort-level. Because 88.1% of named pairs come from a single engine, a company should ask the same buying questions it expects customers to ask, directly, on ChatGPT, Gemini, and Claude separately, using the specific phrasing of commercial-investigation and transactional prompts (the types most directly tied to purchase intent, at 39.1% and 38.3% appearance rates respectively in this study). A company absent on one engine and present on the others has a narrower, cheaper problem than one absent everywhere.

Step two: do not expect position to be the lever. Average position across engines sits close together, around 4.94 out of 197,893 mentions. This suggests that once a company clears the bar to be named, the engines place it similarly. The leverage point is therefore getting named at all, not jockeying for a better slot among the named. Effort spent trying to look more authoritative to an already-favorable engine is probably better spent figuring out why an unfavorable engine does not name the company at all.

Step three: treat category obscurity as a dead explanation. Since visible rates sit at or near full across every category measured, from the smallest (Tourism Digital Marketing, 8 companies) to the largest (E-commerce Solutions, 3,600 companies), a company cannot reasonably conclude its category is too small or too niche for AI assistants to know. The category is covered. The company's own standing within it is the open question, and that is the one worth investigating, through its own web presence, structured listings, and third-party mentions, rather than through its category's size.

Step four: re-measure rather than assume stability. Because this study pins one model version per engine and covers one three-week window, any change a company makes should be checked against a fresh measurement under comparable conditions, not assumed to have worked based on general impression. Model updates outside a company's control can move the numbers independent of anything the company did.

How to read the published dataset yourself

The figures in this study are drawn from a dataset large enough, 536,301 raw rows collapsing to 536,301 usable answers across 58,803 companies, that most readers will want to look up their own category or company rather than take the aggregate figures on faith. A few things to check before drawing conclusions from any single row or cut.

Check the denominator before trusting a percentage. The by-category table in this study ranges from categories with a handful of companies, such as Fitness Digital Marketing at 7 companies or Hospitality Digital Marketing at 9, up to categories with thousands, such as E-commerce Solutions at 3,600 or Real Estate Services at 4,139. A visible_pct figure built from a dozen companies carries far less statistical weight than one built from thousands, even when both render as the identical rate. Treat any small-category figure as a directional signal, not a precise measurement, and widen the lens to a parent category or the full cohort when the count is small.

Check which engine a figure is about. Because engine-level agreement is low, a visibility figure that does not specify which engine it describes is close to meaningless for action. The 95.6% any-engine figure and the 32.4% Gemini-only figure are both true and both in this dataset, and they answer different questions. Know which one you are looking at.

Check the run window and model version. This dataset reflects 2 window boundaries spanning 2026-09-04 to 2026-09-25, against one pinned model version per engine. A figure pulled from this dataset describes that window and those versions, not an enduring state of the world. Model providers change their systems on their own schedule, and a category that looks fully covered today is a statement about September 2026, not a permanent guarantee.

Check whether a result is about presence or about position. The two axes move differently. Being named and being named near the top of the answer are separate measurements in this dataset, and conflating them, treating any mention as equally good, erases a real and actionable distinction. A company appearing once, in a low position, on one engine out of three is a different situation from a company appearing consistently, near the top, on all three, even though both could be summarized loosely as