Do AI Engines Recommend the Same Companies? A 3-Engine Agreement Study

Why this question matters

A buyer researching, say, a blockchain testing vendor or a commercial cleaning service increasingly starts with one AI assistant, not four. They ask ChatGPT, or they ask Claude, or they ask Gemini. They rarely ask all three and compare notes. That means whichever companies a given engine happens to surface become, for that buyer, the entire visible market. If a company's presence in AI-generated answers varies sharply by which engine is asked, then a company's fate in this new channel depends partly on an accident of which assistant the buyer opened.

This matters most to two groups. The first is companies trying to manage their presence in AI answers, who need to know whether optimizing for one engine's behavior helps or does nothing for the others. The second is anyone buying or building an 'AI visibility' monitoring product, who needs to know whether a single-engine score is a reasonable proxy for the whole market or a measurement of one-third of it.

The question is also a check on a common assumption: that large language models, trained on overlapping web corpora and asked the same question, will converge on a similar shortlist of companies the way three independent analysts reading the same industry reports might converge on the same handful of leaders. If that assumption held, a single engine's answer would be a reasonable stand-in for 'what AI recommends.' If it does not hold, the entire idea of a portable, engine-agnostic AI visibility score needs to be qualified with which engine, or engines, it is actually measuring.

This study does not ask whether the companies named are good companies, or whether being named causes more business. It asks a narrower, checkable question: given the same prompt, do three engines converge on the same names. The answer, measured across 30,564 companies and 170,457 answers, is that they mostly do not, and the rest of this study describes the shape of that disagreement in more usable terms than 'mostly.'

How the measurement works

The design starts from a fixed prompt set, 2,500 distinct prompts covering commercial-investigation, informational, navigational, and transactional intents, spanning categories such as SaaS, Legal Services, Real Estate Investment, and Manufacturing Services. Each prompt was sent to all three engines, ChatGPT, Claude, and Gemini, using a single model version per engine (1 version tracked for ChatGPT, 1 for Claude, 1 for Gemini), so differences in results reflect differences between engines rather than between model generations within an engine.

Runs were collected across a measurement window bounded by two timestamps in 2026, and 3 mitigations were applied during collection to reduce known sources of noise, such as retrying failed calls and normalizing company names before matching. The raw log contains 170,457 rows; after filtering to rows with a clean, parseable answer, 170,457 remained, all logged with an 'ok' status (170,457 of them), meaning none were dropped for parsing failure in the final count.

For every usable answer, the study recorded which companies were named and, when named, the ordinal position of the mention within the answer (first company mentioned, second, and so on). This produced two separate measurements that are easy to conflate but answer different questions. The first is whether a company appears at all on a given engine for a given prompt, which is the basis for the agreement analysis. The second is where in the answer it appears when it does, which is the basis for position analysis such as 5.18 as the overall average position across 65,241 scored mentions.

The unit of comparison for agreement is the company-prompt pair: for each prompt, which companies did each engine name, and how many of the three engines named each one. This is deliberately stricter than asking 'did the engines cover similar topics,' because it requires exact company identity to match, not just category or theme. A pair counts as agreed only if the same company name, after normalization, was produced by more than one engine for the same prompt.

The headline result, and its limits

Across 56,819 company-prompt pairs where at least one engine named a company, the breakdown by number of agreeing engines is:

The plain reading is that sole mentions dominate. 88.4% of pairs were named by exactly one engine, and full three-way agreement happened on only 3.2% of pairs.

What this does not mean is that the engines are randomly guessing or that most companies are invisible. Separately, 0% of the 30,564-company cohort never appeared on any engine at all, meaning the disagreement is concentrated among companies that do get named somewhere, just not everywhere. And 96% of companies were visible on at least one engine without being visible on all three, which is the group most exposed to the choice-of-engine problem described in the first section.

It also does not mean the engines are working from different facts about the world. A more likely explanation, consistent with the pattern, is that each engine draws from a differently weighted slice of its training and retrieval sources and applies its own shortlist logic, so two engines can both have accurate information about a company and still not both choose to name it in a given answer, especially in categories with many plausible candidates. Distinguishing 'the engine doesn't know about this company' from 'the engine knows but didn't pick it' would require probing each engine directly about specific companies, which this study did not do; it observed only what engines volunteered in response to prompts a buyer would plausibly write.

A second limit: the study measures naming, not ranking quality or accuracy. A company could be named correctly by one engine and correctly omitted by another because the other engine judged it a weaker fit for that specific prompt. Low agreement is consistent with both 'engines see different parts of the market' and 'engines make different but equally defensible editorial choices.' This study's data cannot fully separate those two stories, though the category-level consistency described later leans toward the former.

Is this a gradient or a cliff

The shape of a result changes what follows from it. If visibility were binary, either a company is known to AI engines or it is not, the practical response would be simple: get past the threshold. If it is a gradient, the response has to be about degree and about which engine, not about crossing a single line.

The data support a gradient. The three-way split in the previous section, roughly 88.4% sole mentions, 8.4% two-engine mentions, and 3.2% three-engine mentions, is not a spike at zero and a spike at three with nothing in between. There is a real, populated middle tier of pairs named by exactly two engines. And per-engine visibility rates form their own gradient rather than clustering at one value: 42.7% on ChatGPT, 39.9% on Claude, 33.6% on Gemini.

The near-zero rate of complete invisibility reinforces this. 0% of the cohort's 30,564 companies never appeared on any engine, which means the practical question for almost every company in this dataset is not 'am I visible at all' but 'am I visible on the engine this particular buyer happens to be using.' That is a meaningfully different problem to solve. A binary visibility problem has a single fix: produce more of the content that gets a company noticed at all. A gradient, engine-specific problem means a company can be doing everything right for one engine's retrieval and ranking behavior and still be absent from another's answer to the identical question.

This also means a company cannot infer its standing on Gemini from its standing on ChatGPT. The position data adds a secondary nuance: once named, companies land in a similar spot regardless of engine, averaging 5.25 on ChatGPT, 5.38 on Claude, and 4.85 on Gemini, against an overall average of 5.18. So the gradient is almost entirely about the yes/no decision to name a company, not about how prominently it gets named once an engine decides to include it.

What the engines appear to be working from

This study did not log source citations for this dataset (no per-source citation data was captured), so it cannot say directly which pages, directories, or reviews each engine drew on when deciding whom to name. What it can say is inferred from the pattern of answers themselves.

Coverage was effectively complete at the category level: across every category examined, from small ones like Healthcare IT Services (9 companies) and Fintech Testing Services (24 companies) to large ones like SaaS (994 companies) and Food Service Business (819 companies), the visible share of companies sits at 100% of companies with at least one visible-somewhere mention in every category checked. That uniform, near-total category coverage suggests the engines are not simply blind to entire industries or company sizes; the disagreement documented above is happening inside categories, between specific competing companies, not between categories.

Prompt intent also shapes how often any company gets named, though less than might be expected. Appearance rates by prompt type are close to each other: 38.6% for commercial-investigation prompts (of 21,885 answers), 38.7% for informational prompts, 36.9% for navigational prompts, and 38.3% for transactional prompts, by far the largest bucket with 124,338 answers and 47,628 appearances. The similarity across intents suggests that whether a company gets named is driven less by what kind of question was asked and more by which specific companies each engine's underlying sources happen to surface for that topic.

The honest limit here is that 'what the engine is reading' is an inference from output patterns, not a direct observation of retrieval logs or training data. A company that wants to know why it was named by one engine and not another cannot get that answer from this dataset; it can only see that the split exists and how large it is.

Where the engines disagree, and why a single score misleads

The practical consequence of the agreement numbers is aimed squarely at anyone buying or building a single 'AI visibility score.' If 88.4% of company-prompt pairs are named by only one engine, then a monitoring tool that checks only ChatGPT, for instance, is not a slightly-incomplete version of the full picture. It is closer to a one-third sample dressed up as the whole.

The per-engine visibility rates make the size of that gap concrete: 42.7% of companies appear on ChatGPT, 39.9% on Claude, and 33.6% on Gemini. A company sitting at, say, strong ChatGPT visibility and weak Gemini visibility would score very differently depending on which single engine a monitoring product happened to track, even though nothing about the company changed between the two measurements.

This is not simply engines disagreeing at random. The five companies that showed the strongest presence across the dataset, The Home Edit, Junk King, BMM Testlabs, Ecobee, and Visit El Paso, a set of 5 companies spanning organizing services, junk removal, gaming testing, smart home devices, and tourism, span very different categories, which suggests that broad visibility is achievable in different market types, not confined to one kind of business or one kind of engine's apparent preference.

What this means for a buyer of an AI visibility product: ask which engine, or engines, the score is built from, and treat a single-engine score as a measurement of that engine's answer behavior, not of 'AI' as a whole. A score built from all three engines, weighted or reported separately, will track the 3.2% of pairs where engines agree and the much larger share where they do not, rather than collapsing both into one number that hides the disagreement. The alternative explanation worth naming honestly: some of this disagreement may reflect legitimate differences in what counts as a good answer for a given prompt, not just gaps in each engine's knowledge. Nothing in this dataset separates 'engine doesn't know about this company' from 'engine knows but chose a different answer,' and a company investigating its own case would need engine-specific probing to tell the two apart.

What a company on the wrong side of this should do

A company that finds itself named by one engine and not the others has a narrower, more tractable problem than 'improve AI visibility' suggests. The data point to a few concrete responses.

First, check per-engine visibility separately rather than accepting a single blended score. Given that 42.7%, 39.9%, and 33.6% are visibility rates that move independently, a company should know its standing on each engine, not just an average across them. An average of a high and a low score looks identical to a flat medium score, and the fix for those two situations is different.

Second, treat the 0% near-zero rate of total invisibility as evidence that the floor is not the problem for almost any company in a covered category. The realistic goal for most companies is not 'get discovered by AI at all,' which is already happening; it is 'close the gap between the engine where I appear and the engines where I don't.' That reframes the work from broad content production toward figuring out what is different about the sources or signals feeding the engine where the company is absent.

Third, do not assume position work matters more than naming work. Because average mention position is close across engines (5.18 overall), a company already being named is not obviously being buried by one engine relative to another. The bigger lever is the binary naming decision itself, not incremental position improvement once named.

Fourth, watch category context. Since visible coverage is at or near total in every category examined, a company that is absent from an engine's answers is not absent because the category itself is under-covered; it is competing against specific named alternatives within a well-covered category. That argues for competitive analysis, comparing against the companies that do get named for the same prompts, over generic visibility campaigns.

Finally, recognize the limits of self-diagnosis. Because this study cannot see each engine's underlying sources, a company cannot fully reverse-engineer why one engine omits it. What it can do is monitor consistently across all three engines over time and treat any engine-specific gap as a hypothesis to test with targeted content and outreach, then remeasure.

How to read the published dataset yourself

The underlying dataset is built from 170,457 logged rows, narrowed to 170,457 usable answers, all with an 'ok' status, collected across a window bounded by two UTC timestamps in 2026. Each row ties one prompt, one engine, and one answer together, with company mentions and their positions extracted from the answer text.

Three tables matter most for independent checking. The first is the pair agreement table, which groups company-prompt pairs by how many of the 3 engines named them; use the pairs count columns, not just the percentages, if you want to know how thin or thick a slice you are looking at, since 1,812 pairs is a much smaller base than 50,209 pairs even though both get summarized as percentages.

The second is the per-category breakdown, useful for anyone who only cares about their own vertical. Categories range enormously in size, from Healthcare IT Services at 9 companies to Digital Marketing at 827 companies, and a category-level visible rate calculated on a nine-company base carries far less weight than the same rate on an eight-hundred-company base, even when both render as the same percentage.

The third is the per-engine visibility and position tables, which should always be read side by side rather than blended, given the disagreement documented in this study. If you build your own summary statistic from this data, report it per engine first, and only then consider whether an aggregate is meaningful for your purpose.

One caution for reuse: this dataset used a single model version per engine (1 for ChatGPT, 1 for Claude, 1 for Gemini), captured across a fixed window. Engine behavior changes as providers update models, so a rerun months later on updated model versions should be expected to shift these numbers, possibly substantially, and should not be treated as invalidating this snapshot or as directly comparable without checking model versions on both sides.