Agencies that get named in ChatGPT and Perplexity, measured

Between 22:30 on 26 July and 01:15 on 27 July 2026 we put ten buyer questions about AI search visibility agencies to Claude, ChatGPT, Perplexity and Google AI Overviews, five times over. Forty answers per run, 200 in total. Six of the ten questions named no agency, which leaves 120 answers where the engine had to pick a name unprompted.

Those 120 answers are the ones worth reading. The other 80 came from questions shaped as alternatives-to or one firm versus another, which hand the engine a name and then reward it for repeating it.

Five agencies accounted for 19 namings across 120 name-free answers. One firm took 15 of the 19.

The question people actually type is engine-specific. They want to know who wins in ChatGPT, or who wins in Perplexity. We are not going to answer that from this sample, and it is worth being direct about why.

Ten questions, five runs, four engines is 50 answers per engine. At that size the difference between an agency being named three times and being named once is inside the run-to-run variation we can already see in the pooled data. Publishing a per-engine league table from 50 answers would look more precise than the measurement is. A pooled count you can trust beats a split you cannot. If you need the per-engine split for your own category, that is what the full audit produces, at ten questions written for your category rather than a shared ten.

One thing did hold across all four engines: the same small group of names kept appearing, and the great majority of answers named at most two of them. The engines are not disagreeing about who exists. They are drawing from a narrow shared pool.

The 200 answers returned 663 citations across 223 distinct domains. 129 of those domains were cited exactly once. That is a long tail carrying almost nothing, and a short head carrying most of the weight.

Every domain in the top group is owned by the agency it represents. The engines are reading vendor sites and summarising them.

Read the two tables together and the useful finding falls out. rampiq.agency was opened 25 times. Rampiq was named in zero of the 120 name-free answers. zupo.co was opened 19 times. Zupo was named in zero of them. Being read is not being recommended, and a citation count on its own will tell you a story your pipeline does not agree with.

If your own category behaves like this one, and most B2B software categories we have measured do, three things follow.

First, your own site is table stakes rather than an edge. Vendor domains take the top of the citation list in this category, which means everyone is already being read. Second, the gap between citation and naming is where the work is. Third, the number that predicts revenue is the name-free count, not the count from questions that already contained your name.

We publish our own numbers on the same basis. Broadcastwell scored 0 of 40 when we ran our four-engine method on ourselves, and that is on our own visibility page.