Same day, same six businesses: one AI engine named three of them, another named none
Why this exists. Every AI visibility tool on the market will sell a business a single number. This is what happened when we asked three different engines the same buyer questions about the same six businesses on the same day. The engines did not agree. They did not disagree a little. On three of the six businesses, one engine named them every single time and another engine never named them at all. A single score cannot survive that, and this is the receipt.
The measurement
On 2026-08-05 we ran six real local businesses through the clean API channel. Not a logged-in browser session, because a logged-in session is personalized and a personalized answer is not a measurement. Each business got one buyer-shaped question, written the way a customer would actually ask it, and none of the questions named the business. That is the point: we are asking whether the engine surfaces them unprompted.
| Business | The question the engine was asked |
|---|---|
| BBS Insurance | What is the best health insurance broker in Kansas City MO for someone who is self-employed? |
| Flatline Contracting | Who is a reliable contractor in Joplin MO for a home remodel or repair project? |
| Paint Saint Louis LLC | Who is the best interior painting company in St. Louis MO? |
| World Martial Arts Academy | What is a good martial arts school for kids in St. Peters MO? |
| Backflow Boss | Who does backflow testing and repair near Liberty MO? |
| Raven Moon Academy | What are the best online courses for learning shadow work and spiritual practice? |
Finding 1, the gap
| Engine | Successful runs | Businesses named at least once | Rows where the business was named |
|---|---|---|---|
| Perplexity | 30 of 30 | 3 of 6 | 15 of 30 |
| Gemini | 16 of 30 | 1 of 6 | 1 of 16 |
| Anthropic | 12 of 12 | 0 of 6 | 0 of 12 |
Read the denominators before the headline. The three engines did not get the same number of runs. Perplexity completed 5 runs per business, Anthropic completed 2, and Gemini completed a partial set because its free tier refused 14 of 30 requests on quota. A refusal is not a failure to answer and we do not count it as one; those 14 rows are labeled and excluded.
Within those denominators the result is stark. Perplexity named BBS Insurance, World Martial Arts Academy and Backflow Boss on all five of its runs each, and named the other three businesses on none of them. Anthropic's saved answers did not name any of the six, on either run (0 of 12 answers). Recounted on 2026-10-12 from the saved answer text by whole-name substring match: Perplexity 15 of 30 rows, Gemini 1 of 16 rows, Anthropic 0 of 12 rows.
What that means for three real business owners. BBS Insurance, World Martial Arts Academy and Backflow Boss were, on 2026-08-05, perfectly visible on one major AI engine and completely invisible on another. If they had bought a visibility score that day, the number they received would have depended entirely on which engine the vendor happened to ask. Not on their business. Not on their website. On the vendor's choice of engine.
Finding 2, what the silent engine said instead
Anthropic's zero is not an error state. All 12 rows returned run_status OK with real answers between 915 and 1,379 characters. The engine answered the question. None of the 12 answers named any of the six businesses. Two of them, for the shadow work question, did name other course providers (Insight Timer in both, The Shift Network in one). Counting across those 12 rows:
- 9 of 12 answers pointed the customer to a review or comparison platform instead of a business. Yelp appeared in 7, Better Business Bureau in 6, Google Reviews in 5, Angie's List in 2, HomeAdvisor in 1, Thumbtack in 1.
- 6 of 12 contained an explicit statement that it could not name a specific option, in phrasings like not having access to current business ratings, or not being able to say definitively which is best.
- 9 of 12 ended by asking the customer a question rather than answering theirs.
Counted by case-insensitive substring match across the stored answer text of all 12 rows.
Here is the shortest of the twelve in full, unedited, from the interior painting question:
The practical read for a local business. When the engine will not name anyone, it hands the customer to a directory. On that query, on that day, Yelp and the Better Business Bureau were the competition, not the painter down the road.
Finding 3, empty cited-sources field
Across all 26 Anthropic rows collected that day, including the segment queries described below, the recorded cited-sources field was empty every time. Perplexity's rows carry source URLs in that field and Anthropic's carried none. This is what the stored field shows. It is not a recount of sources named in the answer text, and stored fields can under-count what the text contains (method note DQ-5, 2026-10-12).
This matters for anyone trying to influence an engine. On an engine that cites, you can see what it read and go work on those sources. On an engine that cites nothing, there is no visible surface to work on at all. Any vendor selling you a fix list for an engine that publishes no citations should be asked how they know.
The correction we caught before publishing
Our own internal note for this finding originally read "named zero of 13 businesses." That sentence was wrong and we nearly shipped it.
The roster holds 13 active rows, but they are not 13 comparable businesses. Six are category research queries with no subject business at all, asking things like how to start a rental arbitrage business. In those rows the field that records whether the business was named has nothing to point at, so a false value there measures nothing. One more row is Namebeam itself, which belongs in its own bucket, not in a customer sample.
The denominator is 6 businesses and 12 rows, not 13. The corrected number is smaller and less dramatic than the one we almost printed. We are printing the near-miss because a company that sells a dated record has to show its own denominators being fixed, not just its findings being right.
Weaknesses, stated plainly
- Six businesses is a small sample. Six is enough to demonstrate that the gap exists. It is not enough to estimate how common it is, and we are not claiming a rate.
- One day. All of this is 2026-08-05. Engines change.
- Unequal runs. Perplexity got 5 runs, Anthropic 2, Gemini a quota-limited partial. A third Anthropic run could name someone and would be published if it did.
- One question per business. A different phrasing could produce a different answer. Our earlier work found the wording of the customer's question matters a great deal.
- Five of the six questions are geographically anchored local searches. One is a national category question. They are not a uniform query type.
- One engine's behavior on one prompt style is not that engine's behavior everywhere. Do not read this as a ranking of engines. It is a measurement of disagreement between them.
- Answers are not stable run to run. In earlier dated work on this corpus, zero of 22 repeat answers were byte-identical. Notably, the naming verdicts here were stable even though the wording was not: Perplexity's three named businesses were named on all five runs, and its three unnamed businesses on none.
Method
- Channel: API only, through an automated runner. Never a logged-in browser session, because personalization contaminates the result.
- Every answer stored verbatim and dated in an append-only public corpus. Failed and quota-refused rows are labeled and kept, never deleted.
- Naming determined by substring matching against documented name variants of each business, not by an entity extractor, so the result does not inherit an extractor's bias.
- Quota refusals are recorded as refusals and excluded from every denominator. They are never counted as an engine failing to answer.
Check our work. Every transcript behind this page is public and dated. The rows for this measurement are the files named NB-CZ-API_2026-08-05 in the corpus at github.com/mrcrtr1979-droid/namebeam-visibility-index. Read them, recount them, and tell us if we got it wrong.
Run the free check on your own business
No card and no account. You get the dated transcripts, not a score.