Four AI assistants, the same question. They named different companies.
We asked ChatGPT, Claude, Gemini and Perplexity the buying questions a customer would ask in eleven industries, without naming a single company. Of the 48 S&P 100 companies that any assistant named, only 32 were named by all four — 67% agreement.
JPMorgan Chase was named by Gemini and by none of the other three. IBM was named by ChatGPT alone. A company checking one assistant would reach the opposite conclusion of a company checking another.
What is Model Agreement?
Model Agreement is the share of companies visible in one AI assistant that are also visible in all the others. We measure it by asking four assistants the same unbranded buying questions and comparing which company names each one returns. Low agreement means a single tool misreads your visibility.
In this study Model Agreement was 67%. The practical consequence: a visibility number taken from one assistant describes that assistant, not your market.
Which companies were visible in only one assistant?
Seven of them. Five were named by Gemini alone and two by ChatGPT alone, across the same questions every assistant received.
| Company | Named only by |
|---|---|
| JPMorgan Chase | Gemini |
| Morgan Stanley | Gemini |
| BlackRock | Gemini |
| Booking Holdings | Gemini |
| Cisco | Gemini |
| IBM | ChatGPT |
| Merck | ChatGPT |
How many companies did each assistant name?
| Gemini | 46 |
|---|---|
| Claude | 39 |
| ChatGPT | 36 |
| Perplexity | 34 |
How many companies does an AI answer actually name?
A handful. In four of the eight industries we could analyse, three companies took every mention — there was no fourth place. In the rest, the top three took between 50% and 75%.
| Industry | Top 3 share | Who |
|---|---|---|
| Energy | 100% | ExxonMobil, Chevron, ConocoPhillips |
| Utilities | 100% | Duke Energy, NextEra, Southern Company |
| Consumer Discretionary | 75% | Amazon, Home Depot, Tesla |
| Industrials | 64% | GE Aerospace, UPS, FedEx |
| Consumer Staples | 62% | PepsiCo, Coca-Cola, Procter & Gamble |
| Communication Services | 50% | Alphabet, Comcast, Meta |
How was this measured?
Four unbranded buying questions per industry, asked 3 times to each of four assistants, then matched against the S&P 100 constituent list. 398 answers were analysed.
- 01
Universe
The 100 constituents of the S&P 100, with their GICS sector, from the public index listing.
- 02
Questions
Four buying questions per industry. No company is named in any question — the assistant volunteers the names or it does not.
- 03
Assistants
One current, provider-hosted model per family: openai/gpt-5-nano, anthropic/claude-haiku-4.5, google/gemini-2.5-flash-lite, perplexity/sonar. Open-weight stand-ins were rejected: they are not what these products run.
- 04
Repeats
Each question asked 3 times per assistant, because these systems are stochastic and a single run is an anecdote.
- 05
Matching
Word-boundary match on each legal name plus curated aliases. Companies whose name is an ordinary English word are flagged in the dataset so their counts can be read with suspicion.
- 06
Exclusions
Financials, Health Care, Information Technology fell below 80% usable answers and are excluded rather than reported thin.
What this study does not show
Four separate defects in our own tooling turned up while running this, and each one had moved a number in the direction that flattered our argument. That is the ordinary direction of measurement error. It is why the script and the raw answers are published with the figures.
- Three sectors are missing
- Financials, Health Care and Information Technology fell below 80% usable answers because of a defect in our own job queue, so they are excluded rather than reported thin. A re-run is in progress.
- We do not publish a 'never mentioned' count
- Our questions were written per GICS sector, and a sector is broader than a company's actual category. Caterpillar makes construction equipment; none of our four Industrials questions asked about it. A company can look absent because its category was never raised, which is a flaw in the question set, not a finding about the company.
- Only one of the four searched the live web
- Perplexity's sonar browses; the other three answered from what they have learned. This measures learned association, not what a consumer product with browsing enabled would return today.
- Being named is not being recommended
- A mention is the start of a shortlist, not a ranking, a click or a sale. We count mentions because that is what we can observe, and we do not extrapolate from them to revenue.
We can run this for your company instead of the S&P 100.
Same method, your category, your competitors. You get every prompt, every answer and every model ID, and you keep the report whether or not we work together.
Show me what AI says about us