The State of AI Search, August 2026

Why this matters, and to whom

When someone asks ChatGPT, Gemini, or Claude which vendor to use for a given job, the assistant answers in prose, not in a list of ten blue links. There is no search results page to scan, no second page to click through to. There is just whatever the model said, usually naming a small handful of companies. If your company is not one of the names it said, you do not lose a ranking position, you lose the conversation entirely.

That is a different kind of exposure than traditional search, and it is why BusySeed ran this study. The question we wanted answered was plain: across a realistic spread of buyer questions, how often do these three assistants mention a given company at all, where in the answer do they tend to put it, and do the three engines even agree with each other about who belongs in the answer?

The audience for this is anyone whose revenue depends on being the name a model says out loud: marketing teams who have historically optimized for search rankings, agencies now being asked to explain "AI visibility" to clients, and the clients themselves who are being sold dashboards claiming to score it. All three groups need a baseline that was not produced by the vendor selling the dashboard.

It also matters because the underlying behavior is new enough that intuitions from search engine optimization do not transfer cleanly. A page can rank on Google through a mechanism anyone can inspect: backlinks, keywords, crawl status. What an assistant chooses to say in response to a prompt is the output of a language model synthesizing training data, retrieved context, and whatever ranking logic sits behind that particular product, and that process is not something a company can audit from the outside. The best available move is to measure the outputs at scale and look for patterns, which is what this study does: 100,966 companies, 14,951 distinct prompts, run against all three engines across the four-week window of August 2026.

The finding that should concern a reader most is not the headline visibility percentage. It is the disagreement rate between engines, covered later in this study, because it determines whether "AI visibility" can be a single number at all, or whether it has to be reported per engine, the way search visibility eventually had to be reported per region and per device.

How the measurement works

The design goal was to approximate what a real prospective buyer does: ask a plain-language question and see what an assistant volunteers, without logging into an API playground or crafting a prompt engineered to surface a particular brand.

The cohort. 100,966 companies were tracked across categories ranging from narrow ones like Legal Digital Marketing (10 companies) to broad ones like Mobile App Development (6,048 companies). Category size varies by two orders of magnitude in this dataset, which matters when reading any category-level cut: a percentage computed over companies in a category like Tourism Digital Marketing (3 companies) is a statement about a much smaller sample than the same percentage computed over Education IT Services (4,049 companies), and should carry proportionally less weight.

The prompts. 14,951 distinct prompts were built to mimic buyer-style questions across informational, navigational, commercial-investigation, and transactional intents, the same four-way split search researchers have long used to classify query intent. Each prompt was run against each engine inside the study window, August 4 through August 31, 2026, producing 1,013,625 total rows.

The engines. Three assistants, 3, each queried on a single model version held constant for the duration of the run (1 version for ChatGPT, 1 for Gemini, 1 for Claude). Holding the version constant matters because these products update frequently and silently; a study that spans a version change cannot tell you whether a shift in results is real or an artifact of the vendor shipping a new model mid-study.

What counts as usable. Of 1,013,625 rows produced, 1,013,625 returned a valid, parseable answer, a status rate of 1,013,625 usable answers out of 1,013,625 attempted. A row fails parsing if the engine declines to answer, times out, or returns a format the extraction pipeline cannot read as a ranked list of company mentions. The clean status rate here means the usable-answer figures in this study are not quietly built on a smaller, unrepresentative subset.

This design cannot see what the assistant would have said to a different phrasing of the same question, or to a follow-up question in a multi-turn conversation. It measures first-answer behavior to a single well-formed prompt, which is a meaningful but partial slice of how people actually use these tools.

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. Every answer came from the vendor's small, fast model tier responding from the model's own knowledge, with web search switched off. Where the model name was recorded, it was OpenAI's GPT-4o mini, Anthropic's Claude Haiku and Google's Gemini Flash Lite. Rankxa began stamping the model name on each answer partway through August, so earlier answers in this window carry no stamp, although the configuration did not otherwise change. Read these results as what each vendor's fast tier recommends when asked cold, which is the tier most third party tools build on, not as what a logged in consumer sees.

The headline result, and its limits

Across the full run, 1,013,625 usable answers were collected. Within those, companies appeared at an average ranked position of 4.95, computed over 363,777 individual mentions. No company in the cohort was completely invisible: 0 companies, 0% of the cohort, never appeared on any engine for any prompt.

That last figure is the one most likely to be misread, so it is worth being precise about what it does and does not claim. It means every company in this cohort was named by at least one engine on at least one prompt somewhere across the month. It does not mean every company was easy to find. A company might owe its sole appearance to one prompt, on one engine, at the bottom of a long list. The measure that distinguishes "technically visible" from "actually competitive" is how many companies were visible somewhere but not everywhere: 96,644 companies, 95.7% of the cohort, showed up on at least one engine without showing up on all three.

Put those two numbers side by side and the shape of the result becomes clear. Universal invisibility is effectively a non-issue in this cohort. Universal visibility, meaning a company that appears consistently across all three engines, is the rare case, not the norm. The vast majority of companies occupy a middle zone: visible to some engines, for some prompts, and absent elsewhere.

This is also where the by-engine appearance rates matter. ChatGPT showed 43.9% of companies at least once, Claude 40.9%, Gemini 32.6%. These are three different populations of "visible companies," not the same population measured three times. A company can read its own near-zero appearance rate on one engine next to a strong one on another and conclude, correctly, that the problem is not a lack of digital presence generally, it is something specific to how one engine's underlying retrieval or training data treats that company.

Gradual, not binary, and why that follows from the position data

A natural first guess about AI visibility is that it behaves like a light switch: either the model knows your company and says so, or it does not. The data does not support that model. Visibility here behaves more like a dimmer, with gradations in both whether a company appears and where it lands when it does.

Evidence against binary, from presence. If visibility were binary, the never-appears figure and the appears-everywhere figure would account for nearly the whole cohort, with little left over. Instead the bulk of companies, 96,644 of them (95.7%), sit in between: present on some engine-prompt combinations and absent from others. That is the signature of a gradual effect, something closer to a probability of being mentioned on any given prompt than a fixed yes-or-no trait.

Evidence against binary, from position. Among the mentions that do occur, position varies by engine rather than clustering at one spot. Average position is 4.95 overall, but it ranges from 4.72 on Gemini, the earliest of the three, to 5.08 on ChatGPT, the latest, with Claude in between at 5. A company does not simply "make the list" or not. It makes the list at a position that an engine-specific tendency shapes, on top of whatever is specific to that company and that prompt.

What follows from gradualness. If the result were binary, the right response would be a one-time fix: get your company into the training data or the retrieval index, and the problem is solved. Because it is gradual, the more accurate framing is a continuous optimization problem, closer to how companies have long treated organic search ranking than how they treat, say, being listed or not listed in a directory. A company sitting at an average position past the midpoint of typical answers is not locked out; it is a candidate for the kind of incremental work, covered later in this study, that shifts position and appearance rate over time rather than flipping a switch.

The practical implication is that anyone tracking this should expect to move a percentage gradually across months, not to cross a threshold once and be done. A single month's snapshot, like this one, is a position in that gradual movement, not a verdict.

What the engines are reading when they decide who to recommend

None of the three engines in this study expose their retrieval process, so this section states plainly what the method can and cannot see on that question.

What the data can establish is behavioral, not mechanical: how often each engine mentions companies, and roughly where it places them, by prompt type. The four intent categories tracked here show a narrow but real spread in how often an answer contains a company mention at all:

The rates sit close together, within a few percentage points of each other across all four intents, which is itself informative: it argues against the idea that these assistants reserve company recommendations for obviously commercial questions and stay neutral elsewhere. An informational question gets a named company in the answer at close to the same rate as a transactional one.

What the data cannot establish is why a specific company gets named. This study did not capture citation lists or source URLs the engines drew on (no cited-source data was recorded in this run), so it cannot say whether a mention traces to a company's own site, a review aggregator, a news article, or the model's training data with no live retrieval at all. That is a meaningful gap, because the fix for each of those is different: a retrieval-grounded mention can potentially be influenced by what is published and indexed now, while a training-data mention reflects a snapshot that will not move until the underlying model is retrained.

The most defensible working assumption, consistent with how these products are generally understood to operate, is a mixture of both: some retrieval over current web content, blended with whatever the base model already associates with a category or query. The engine disagreement data in the next section is one of the few available ways to reason indirectly about that mixture, since a company known mainly through stable, old training data should behave differently across engines than one known mainly through live retrieval of recent content.

Where the engines disagree, and why one score cannot capture it

If the three engines were drawing on the same underlying knowledge and applying similar logic, they would tend to name the same companies for the same prompts. They do not.

Across 337,500 engine-prompt pairings where at least one engine named a company, the breakdown by how many engines agreed was:

The dominant pattern, by a wide margin, is a sole mention: 88.7% of pairings involved only one engine naming the company at all, with the other two silent on that same prompt. Full three-way agreement happened in only 2.8% of cases. This also matches the full-cohort figure cited earlier: across the whole study, all three engines agreed on a company in only 2.8% of pairings.

This is the finding with the most direct consequence for how this data should be bought and sold. A vendor offering a single "AI visibility score" is necessarily collapsing three distinct, mostly-disagreeing populations of answers into one number. Given that 88.7% of pairings are a single engine acting alone, a composite score mainly reflects whichever engine happens to be weighted most heavily in the formula, which is rarely disclosed. A company could reasonably be strong on one engine and silent on the other two and still receive a mediocre blended score that obscures both the strength and the weakness.

The more honest alternative, and the one this study's own breakdowns argue for, is to report visibility per engine, the same way a business might once have reported organic rankings separately for Google and Bing rather than averaging them. Two explanations are consistent with this level of disagreement: either the engines are drawing on meaningfully different source material (different training cutoffs, different retrieval systems, different licensing deals with content providers), or they are drawing on similar material but applying different selection logic on top of it. Distinguishing those two would require the citation-level data this run did not capture, which is a limitation worth stating plainly rather than papering over with a single blended score.

What a company on the wrong side of this should actually do

A company that checks its own name against this kind of data and finds low or no visibility has a narrower, more specific set of first moves than "do more content marketing."

Separate the engines before doing anything else. Given how rarely the three engines agree, with 88.7% of pairings decided by one engine alone, a company should look at its ChatGPT, Gemini, and Claude results as three separate problems rather than one. Fixing visibility on one engine is not expected to move the others, and a company that only checks one engine before declaring success is checking a third or less of the picture.

Check the basics that plausibly feed retrieval. Since at least part of what these assistants surface is likely grounded in current web content rather than static training data, the fundamentals that have always mattered for being findable and parseable still apply: a clear, specific description of what the company does and for whom, structured and consistent across the company's own site and the third-party pages that describe it (directories, review sites, trade press), and no ambiguity about category. A company whose public description is vague or inconsistent across sources gives a model less to latch onto when it is deciding whom to name for a specific buyer question.

Expect gradual movement, not a threshold. Because the visibility data looks graded rather than binary, as the earlier section on the shape of the result argued, a company should track its own appearance rate and average position over a period of months, not check once and treat a poor result as permanent or a good result as secure. Position moved from 4.72 on the best-placing engine in this study to 5.08 on the worst, which is itself evidence that position is sensitive to something engine-specific and not fixed.

Be skeptical of any tool offering one number. Given the disagreement data above, a vendor product that reports a single AI visibility score without breaking it out by engine is making a modeling choice that this study's own data argues against. The reasonable question to ask that vendor is which engines fed the score and how they were weighted, before trusting it as a diagnostic.

Reading the published dataset yourself

The figures in this study are drawn from a published dataset covering the full run: 100,966 companies, 14,951 prompts, three engines, 1,013,625 total rows, 1,013,625 of them usable. Anyone working with the underlying data directly should keep a few things in mind that are easy to miss from the summary tables alone.

Denominators vary enormously by category. The by-category breakdown ranges from single-digit cohorts, like Tourism Digital Marketing at 3 companies, to cohorts in the thousands, like E-commerce Solutions at 3,296 or Mobile App Development at 6,048. A percentage computed on a category with a handful of companies can swing enormously from a single company's result changing; treat any category-level claim as provisional until you have checked its denominator.

"Visible" is not one threshold. The dataset distinguishes companies that appeared on at least one engine from those that appeared on every engine, and the gap between those two figures, 96,644 companies wide, is where most of the cohort actually sits. Any downstream use of this data should state which definition of visible it is using, because the two tell different stories.

The window is four weeks, once. The run covered August 4 through August 31, 2026, against a single held model version per engine. It is one month's measurement, not a trend line. Comparing this month's figures to a future run is reasonable; treating this single run as a permanent characterization of any one company's standing is not, especially given how these products update.

No citation data in this run. As noted earlier, this run did not capture which sources an engine drew on for a given mention, so the published dataset cannot answer questions about why a specific company appeared, only whether and where. Anyone wanting to test hypotheses about mechanism will need to pair this dataset with their own citation-tracking, logging what each engine references when it answers a given company's category of prompt, repeated across enough prompts to separate a stable pattern from one engine's noise on one day.