Thirteen US Agencies Publish a Ranking of the Best GEO Agencies. All Thirteen Put Themselves First.

“AI SEO agency” is searched 933 times a month in the United States. Its keyword difficulty is zero and its cost per click is $82.43. A term nobody has built authority on, where a single click costs more than a decent lunch.

We looked at who occupies that page, and the page next to it. Counting only what Google itself prints in the result snippets, thirteen distinct US agencies publish a ranking of the best GEO or AI SEO agencies, and each one places itself at the top. We ran the same count in France last week and found eight.

How this was measured: Google organic results, United States and English, 7 August 2026, for the two queries quoted below. We count only what Google displays in its own snippet: where the snippet shows the ranking and its number one, we record it. We did not open all twenty pages, so the figure is a floor rather than a total. The French figures come from the same count run on France on the same day.

The demand, before the rankings

Every volume above has close-variant disaggregation on. The first version of this table did not, on three of the six rows, and it showed 1,000, 590 and 480 where the separated figures are 933, 492 and 360. The difficulty and the cost per click were identical either way, which is what made it easy to miss: only the volume moves. We publish a whole study on that failure, so carrying the grouped number in our own table was not a small thing.

Two things stand out, and they point in different directions.

France has the volume. The United States has the money. The leading French term is searched more often than any single US term here, and the leading US term costs five times more per click. A US agency client is worth dramatically more, or at least US agencies believe so and bid accordingly.

Difficulty is zero almost everywhere. Four of the six terms above have a keyword difficulty of nought. In a category where a click costs eighty dollars, nobody has built the authority to own the term.

That combination, expensive clicks and no established authority, is what produces the page we are about to describe.

The count, United States

Query: “best geo agencies”, United States, English, 7 August 2026. Nine of the first twenty results are rankings whose own Google snippet shows the publisher in first place on its own list:

beomniscient.com, thriveagency.com, qoulomb.com, lawseo.com, directiveconsulting.com, minuttia.com, yesoptimist.com, thedigitalelevator.com and animalz.co.

Query: “ai seo agency”, same market and date, adds four more: searchbloom.com, coalitiontechnologies.com, spicymargarita.co and growandconvert.com.

Thirteen distinct agencies, each first in its own ranking. Because we counted only visible snippets, the real number is higher.

We are not reprinting how any of them phrase it. The finding is the position they give themselves, which anyone can confirm by running the two queries, and their marketing copy adds nothing to it. We sell GEO consultancy in this market, so repeating a competitor’s superlative on our own domain would be doing their distribution for them. That rule cost us a paragraph in the Spanish version of this study before it was written down.

The rankings that name somebody else

It would be unfair to imply nobody does this honestly, and the exceptions are the interesting part.

• designrush.com is a directory. It ranks Funnel Boost Media, WebFX and others, and sells no GEO services itself.
• nutshell.com is a CRM company. It ranks WebFX first and has no agency to promote.
• mikekhorev.com names Nine Peaks Media.
• ziptie.dev, a tool rather than an agency, names Onely.

The pattern is not that honest rankings are rare. It is that the honest ones are written by people who are not competing. A directory, a CRM vendor, a tooling company and an independent consultant have no seat to win, so their lists read differently.

That is the same conclusion our Spanish measurement reached from the other direction: when we asked AI Overview who to hire for GEO in Spain, the stable citations went to third-party rankings and sector press, not to the agencies’ own lists.

Where the buyers actually go

The most quietly damning result is not a ranking at all.

Reddit is the third organic result on both US queries. On “best geo agencies” it is a thread asking which agencies specialise in LLM optimisation. On “ai seo agency” it is a thread whose top comment recommends a firm for optimising brand mentions in AI shortlists.

When a page of rankings is written by contestants, buyers route around it and ask people. That is not a new phenomenon in B2B services, but it has a specific consequence in this category: AI answers cite Reddit heavily, and it appeared in the stable set of several of our own measurements across markets. A thread you did not write may be doing more for your visibility than the ranking you did.

Three markets, one genre

Each term is named because a volume without its keyword cannot be checked, and all three were pulled with market and language set explicitly and close-variant disaggregation on. All three names have other meanings, so we read each SERP as well: GEO agencies own all three.

The genre is identical in both markets we counted, and the United States has more of it. That is what you would expect from the CPC: the more a client is worth, the more worthwhile it is to publish a list with yourself on top.

We later asked the assistants directly in the United States too, and two of four opened with the agency that published the ranking they were citing. One of those sources is in the table above: the self-ranking counted here as a SERP result came back hours later as a citation inside an AI answer, with its publisher named first.

Spain we approached differently, by asking the AI rather than reading the SERP, and it produced the complementary finding: across fourteen runs and four phrasings, forty-nine domains were cited and not one appeared in every run. A market with no settled authority, measured from the answer side instead of the results side.

The fourth market, where we opened the pages

The first bullet in the limits below has been sitting there since 7 August: we counted what Google prints in its snippet, so thirteen is a floor. On 19 August we ran the same question in Germany and opened every page.

geo agentur is searched 1,000 times a month in Germany at a keyword difficulty of 5, and the results page is our category rather than the other meaning of those three letters: all twenty results are about Generative Engine Optimization, none about geodata.

Five ranking pages could be read. 4 of them are published by someone who sells in the category, and all 4 are in their own list. Three of those four are agencies, and all 3 put themselves first. The fourth is not an agency at all.

The fourth row is what opening the pages bought. finance-seo.de publishes a list of 12 and is in it, and the list is alphabetical, so it lands at 5. From a snippet that is one more list written by somebody who sells the thing being ranked, which it is. What a snippet cannot say is where in the list the publisher put itself, and here the answer is: wherever the alphabet puts it.

That is a different behaviour, not a smaller version of the same one, and our floor method would have recorded it as neither.

The last row is the control. marktforschung.de is a market-research trade portal and sells nothing in this category. It names 9 agencies and itself none. One row is not proof, and it is the cheapest available sign that the shape comes from who writes rather than from the format.

Three more of the twenty results are not written at all. They are placements on large publisher domains, each promoting exactly one agency, and all 3 disclose it in their own words: one carries a line saying the advertising partner supplies and is responsible for the content, one says “Advertorial” seventeen times, and one sits on a newspaper’s press-release subdomain.

One of those nearly went into this section as undisclosed, because the first pass counted a single German word, found it once, and that once was inside a cookie notice. The page says the other word seventeen times. The full German write-up, with the two queries verbatim, is here.

And then we opened the American, French and Spanish ones

The German arm was one market. On the same day we ran the two US queries again and opened every ranking on page one, then France, then Spain. Two things change at once and both are said out loud: the instrument, from snippet to opened page, and the date, twelve days after the original count. So this is a fresh reading and not an upgrade of the 7 August number, which keeps its date and its method.

Across the four markets: 36 ranking pages, 29 readable, 7 behind a bot challenge.

21 of the readable pages are published by someone who sells in the category they are ranking. All 21 name themselves. 19 put themselves first.

In the United States alone, where the original count lives, 13 of the readable rankings are published by a seller in the category and 12 of the 13 lead with themselves. The exception is qoulomb.com, which publishes a list of 7 and puts itself at 3, read twice because a surprising cell read once is a reading and not a fact.

7 readable pages are published by someone with nothing to sell in the category. None of them names itself. A CRM company, a rank-tracking tool, a market-research trade portal, a French business podcast, a no-code agency, a Spanish business school and an agency directory each publish a ranking of GEO agencies, and not one of the seven appears in its own list. That is the comparison the finding needs, and it now holds in four markets rather than one.

One page belongs to neither group and is counted on its own. agencegeo.pro is a French domain whose homepage title is the ranking, and its visible text names no publisher: we searched it for the usual French words for a legal notice and an editor and found none. It ranks ten agencies and is none of them.

Spain, which this study had listed as its own gap

Spain is where the limits below used to say we had never counted the results page this way. We had, in Spanish, on 7 August, from snippets, and it gave 4 self-ranking agencies. Opening the pages twelve days later gives 4 again, and 4 of 4 lead with themselves.

The four are not the same four. One of the original four now answers a bot challenge, and one that was not in the earlier count is in this one. A count that reproduces while its membership moves is worth more than either half on its own: it says the shape is stable and the roster is not.

Spain is also the hardest market to read. 4 of the 10 Spanish ranking pages recorded here answer a bot challenge, against three across the other three markets put together, which is why every figure in this section is over readable pages and says so.

Two Spanish shapes are worth naming because the other markets did not produce them. One publisher’s ranking has itself at position 1 and categories of supplier rather than companies at positions 2 to 5, so it is the only named company in its own ranking. And a financial news site carries the whole thing again on its press-release path, attributed to a business school: the same purchased-distribution shape the German arm found on two newspaper domains.

The snippet method is wrong in both directions

This is the part that matters for our own earlier count, and it is uncomfortable.

On 2 of these pages Google’s snippet and the opened page disagree about who the list names first, and they disagree in opposite directions:

• For qoulomb.com the snippet begins by naming Qoulomb ahead of everyone else. The page puts Qoulomb third of seven. A snippet count records a self-ranking that the page does not contain.
• For spicymargarita.co the snippet opens with a different agency entirely. The page puts Spicy Margarita first of nine. A snippet count misses a self-ranking that is there.

So “thirteen is a floor” was the right caveat and not the whole caveat. A snippet count is not a low estimate of the same quantity, it is a different quantity, and it can move either way on any given row. Both of these are in the register with their lists.

We tried to test whether being cited beats being first, and could not

Our Spanish study says, from one market and from the answer side, that appearing in someone else’s ranking is worth more than publishing your own. With 29 classified ranking pages the join looked like one call per market, so we froze two predictions and a base-rate rule and made the calls: the buying question in each market, on AI Overview, five runs each with the cache skipped.

The predictions are not met at a size that means anything, and that is the result.

Two things stop it. The tool returns cited domains and not URLs, so a domain that publishes a ranking can be cited for its services page instead. And two of the Spanish domains that were cited answer a bot challenge, so their publisher was never read on its own homepage, which is the rule this study uses everywhere else. What is left is 7 citations whose publisher kind is known, split 4 sellers to 3 non-sellers, against a population split 21 to 7. Seven observations cannot separate those, and saying otherwise would be reading a rounding error as a finding.

We are writing that up rather than dropping it because the base-rate rule was frozen before the calls, and this is what it is for.

Two things did come out of it, and both are measurements rather than inferences.

The two-word German query returned no AI Overview at all, in 5 of 5 runs. Not an empty answer and not a failed call: the surface did not appear, for a string worth a thousand searches a month at a keyword difficulty of five.

We first wrote that sentence as a fact about Germany, which it is not, and the correction is below because this study is about how a method can be wrong rather than about being right the first time.

One word longer and it renders. Asked again in the same market and language, on the same surface, with the cache skipped: beste geo agentur answers 5 of 5, geo agentur deutschland answers 5 of 5, and the question form answers 4 of the 5 requested. A German query from an unrelated category answers 5 of 5 as the instrument control. So 3 of 3 sibling wordings render and the null belongs to the exact two-word string and to nothing larger.

This register had already recorded the same shape once, in accounting software, where a two-word German head term returned no AI Overview in ten runs while its own longer variants rendered. Two categories, same behaviour, and neither of them a fact about the country.

And the German answer, once it renders, is mostly the pages this study opened. For beste geo agentur, six of the seven domains cited in every run are pages classified above: three of the self-ranking agencies, both of the newspaper domains carrying a disclosed advertorial, and the agency that one of those advertorials promotes. 2 domains are cited in every run of every German wording that renders, and one of them is the newspaper carrying the advertorial.

That is the part a buyer can act on. In this market the most reliably cited source for the buying question is not the best ranking: it is a placement that somebody bought, and the page says so in its own words.

The cited set is stable enough to work with. Spain was read twice at the same bar, and the four domains cited in every run are the same four both times. What differs is the tail: 7 distinct domains in the first reading and 11 in the second. So the core is a real object and the long tail is not, which is what this register has found on every other question it has asked twice.

And one concrete thing, offered as an illustration and not as evidence: the US ranking page cited in all five runs alongside the agencies is published by a CRM company, and that company is not in its own list.

What this means if you are choosing an agency

Read the byline before the ranking. If the page is published by an agency, check whether that agency is number one on it. In the US sample, every agency-published ranking we could read placed its publisher first. That does not make the other nine names on the list wrong, but it tells you what the list is for.

Prefer lists by people with nothing to sell you. Directories, tooling vendors, CRM companies and trade press have no seat to win. They are also, on our Spanish data, the ones AI systems cite.

Ask for the measurement, not the ranking. Any agency claiming AI visibility results should be able to tell you the prompts, the market, the language, the surface, the dates and the number of runs behind its numbers. Our own work this month shows why the last one decides everything: how many runs a category needs varies from three to thirty, and a single measurement in an unstable category is noise with a chart on it.

What this means if you are an agency

Publishing your own ranking is cheap and the page you are competing with is also somebody’s own ranking. Thirteen of them exist in the US alone, they contradict each other, and a buyer who opens two understands the format immediately.

The lists that get cited belong to people outside the contest. Getting into one of those is relationship work rather than content work, and on our data it is what buys the durable citation.

The other lever is the one nobody in this category seems to want: video and forums. Two of the twenty results on “best geo agencies” are YouTube videos, Reddit is third on both queries, and both formats appear repeatedly in the sources AI answers lean on.

What this study does not show

• The 7 August count did not open twenty pages per query. It counted what Google prints, so its thirteen is a floor. The 19 August arm opened them in all four markets and found that a snippet count moves in both directions rather than only downward.
• Ranking yourself first is not misconduct. It is signed, public marketing. We are describing the structure of a results page, not the ethics of its authors.
• Two queries per market, and one day per market. SERPs move. The method is above so you can repeat it, and the German arm was measured twelve days after the other three.
• Nothing about agency quality. Being on a list, or off it, says nothing about whether the work is good.
• The citation arm above tests nothing. It is reported because its design was frozen first and its base-rate rule said, before the calls, that a small intersection could not carry a claim. Read the German null and the Spanish repeat, not the four-to-three split.
• The Spanish column in the table above is from a different method. It reports what the AI answered rather than what the results page contained. The Spanish results page was counted separately, in Spanish, and both of those counts are in the Spain section above.
• CPC is what advertisers bid, not what a client is worth. It is evidence of belief, not of value.

Common Questions About GEO Agency Rankings

Are “best GEO agency” lists trustworthy?

Frequently not. In the United States, counting only what Google displays in its snippets, thirteen distinct agencies publish such a ranking and each places itself first. The lists that name somebody else tend to be published by directories, tool vendors or CRM companies, which have no agency to promote.

How much does a GEO agency click cost?

$82.43 for “ai seo agency” in the United States, and $51.35 for “geo agency”, measured on 7 August 2026. The equivalent French term costs €16.65 and the Spanish one €19.08. The US market pays roughly five times more per click than the French one.

Which market has the most GEO agency demand?

By volume, France: “agence geo” is searched 1,600 times a month, more than any single US term we measured. By value, the United States, where the leading term costs five times more per click. Spain is much smaller at 260 searches a month.

Why is keyword difficulty zero for such expensive terms?

Because nobody has built lasting authority on them yet. Four of the six agency terms we measured across three markets have a difficulty of nought, which means the category is young enough that the pages ranking are recent and lightly linked. That is exactly the condition in which anyone can publish a ranking and appear.

Should my agency publish its own ranking?

The data suggests low returns. There are at least thirteen in the US already, they contradict each other, and the rankings that get cited by AI systems in our Spanish measurement belonged to third parties with no commercial interest in the outcome. Appearing on somebody else’s list is worth more than publishing your own.

Where do buyers go instead?

Forums, on this evidence. Reddit is the third organic result on both US queries we ran, in both cases a thread asking humans for recommendations. Reddit also appears repeatedly among the sources AI answers cite, so a thread you did not write may be shaping your visibility more than the page you did.