GEO agencies, measured

Every ranked list of AI search visibility agencies we can find was written by one of the agencies on it. We did not write another. We ran ten buyer questions across four engines, five times over, scored all 200 answers, and then found a flaw in our own method that changed the result.

200

answers scored

120

of those asked a question naming no agency

of the five agencies appear in none of those 120

663

citations opened

What we ran

Between 22:30 on 26 July and 01:15 on 27 July 2026 we put ten buyer questions to Claude, ChatGPT, Perplexity and Google AI Overviews, and repeated the exercise five times. Forty answers per run, 200 in total. Every answer was scored for which agencies it named, and every citation it returned was opened and classified. The question sets were generated fresh for each run rather than held constant, so figures are not comparable run to run. What follows compares question types instead.

The flaw in our own measurement

Our question generator takes the competitor list as an input. So four of the ten questions in every run had an agency's name inside the question itself: what are the best alternatives to a named firm, or how one named firm compares to another. An answer to a question like that will almost always contain the name that was in the question. Counting it as a mention measures our own prompt, not the engine.

Splitting the 200 answers on that line changes the picture substantially.

Agency	All 200 answers	The 120 that named nobody
Omniscient Digital	49	15
Hamster Garage	19	3
Minuttia	15	1
Rampiq	14	0
Zupo	13	0

Mentions across all 200 answers, against mentions in only the 120 answers whose question named no agency.

Rank order survives the split, which suggests relative position is a real signal. Magnitude does not. The leader's count falls by roughly two thirds, and two of the five agencies appear nowhere at all once the questions stop supplying their names. Everything those two scored came from questions that already contained them.

What the engines read

Rank	Domain	Who publishes it	Citations
1	beomniscient.com	Omniscient Digital	41
2	hamstergarage.com	Hamster Garage	39
3	rampiq.agency	Rampiq	25
4	minuttia.com	Minuttia	24
5	discoveredlabs.com	Discovered Labs	20
6	zupo.co	Zupo	19
7	yesoptimist.com	Yes Optimist	17
8	embarque.io	Embarque	15
9	20northmarketing.com	20North	9
10	piperocket.digital	PipeRocket Digital	9
11	onely.com	Onely	9
12	ekamoira.com	Ekamoira	9
13	demandgenreport.com	Demand Gen Report	9

663 citations across 223 domains. 129 domains were cited exactly once.

Twelve of the thirteen most-cited domains are agencies' own websites. The first source that is not an agency is a trade publication at rank thirteen. There is no analyst firm and no review platform anywhere near the top of this list.

That is the mechanism behind the rankings. When a buyer asks which agency is best at AI search visibility, the engines answer largely by reading what agencies have written about themselves and about each other.

What this means if you are buying

Three things follow from the data, and none of them require taking our word for it.

Ask how the questions were written. Any AI visibility number built from a prompt set that names competitors is measuring the prompt as much as the engine. That includes ours, until we split it. If a vendor reports a share-of-voice figure, ask what is in the denominator.

Being named often is evidence of published presence, not of service quality. The agencies at the top of this table have written more, and been written about more, than the ones below them. That is what the measurement detects. It is not a verdict on the work.

The reference layer is thin. 129 of 223 cited domains appeared exactly once. No independent authority holds this category, which means the position is still open, and it also means every list you read was written by someone with a stake in it.

Method, and our own number

Questions. Ten questions per run, five runs, put to four engines with live web search: Claude, ChatGPT via the OpenAI API, Perplexity sonar-pro and Google AI Overviews. Question sets were generated per run rather than held constant. Four of the ten in each run named a competitor; those 80 answers are reported separately from the 120 that did not.

Scoring. Each answer was scored for whether an agency was named. Brand matching is word-boundary. Every citation returned was opened and its publisher identified by hand.

Our own result. We ran this on ourselves as well. Broadcastwell was named in none of the 200 answers, and broadcastwell.com appears in none of the 663 citations. We are new and have almost no third-party presence, which is the thing this measurement detects. We are publishing it because a firm that measures visibility should show its own number before it shows anyone else's.

Limits. One category, one evening, and 120 answers on the clean split. Because the question sets regenerated between runs, we make no claim about run-to-run stability. Answer-log snippets are truncated, so our own count of zero is a floor rather than a certainty, though it is corroborated by zero citations. We measure presence, not sentiment. Being named is not the same as being recommended well.

Scope. We did not rank these agencies and we are not recommending any of them. Every description above is a count. Anything you read here about what an agency is good at came from that agency, not from us.

Run the same measurement on your own category

Five buyer questions through two engines, results by email in about ten minutes. No call required.

Run your free instant check