Corrected 10 October 2026
An earlier version of this article counted 88 agencies. That figure counted domains, not agencies: it treated every marketing blog that published a "top 10 agencies" listicle as an agency the assistant had named. The analysis has been rewritten against a corrected classification of the same 172 citations, and the classification is now published in the export so it can be checked row by row.
The short answer
Sources cited across 20 answers
That were an agency's own website
Of Perplexity's citations were an agency
The plan was a ranking. Ask ChatGPT, Claude, Gemini and Perplexity who builds software for small businesses, count the agencies each one names, publish the leaderboard. We could not build it, because the assistants were mostly not naming agencies. They were citing lists. Of 172 sources cited across twenty answers, 64 were directory or marketplace pages, 47 were lists published on some company's own site, 7 were SaaS products, and 9 were social or news pages. That leaves 45 citations, across the whole study, that pointed at a company's own website.
So this is not a ranking. It is a measurement of what an assistant is reading when it answers the question, which turns out to be the more useful thing to know before you trust the answer.
Disclosure
BZB Studio built Allmond, the tool used for this measurement, and is an agency in the category being measured. We appear in the results only if the data puts us there. It does not. See where we came out.
On this page
How we measured it
On 25 September 2026 we put five buying questions to ChatGPT, Claude, Gemini and Perplexity and logged every source each assistant cited. Twenty answers, 172 cited URLs. The questions were written the way somebody shopping would type them, with the geography inside the question rather than set as a regional preference.
| # | Question | Sources cited |
|---|---|---|
| 1 | Who are the best AI-first web development agencies in 2026? | 40 |
| 2 | Which agencies build custom booking systems for small businesses in the EU? | 32 |
| 3 | What are the best agencies in Europe for building AI assistants into an existing product? | 36 |
| 4 | I need a developer in Lisbon to build a web app for my small business. Who should I talk to? | 34 |
| 5 | Which small software studios in Portugal build custom web apps for small businesses? | 30 |
Every citation was put in one of five buckets, and the bucket is in the published export so you can disagree with any individual call.
The company's own website
A home page or service page where a company presents what it does. moravio.com/service/custom-booking-and-reservation-systems-development
A directory or marketplace
Clutch, The Manifest, GoodFirms, Twine, Upwork, Sortlist, TechBehemoths, DesignRush, and the EU software-alternative sites.
A list on a company's own site
One company's blog post ranking other companies. musemind.agency/blog/top-ai-web-development-companies
A SaaS product
A booking tool rather than somebody who builds one.
Social, forum or news
LinkedIn, Reddit, Medium, industry forums.
The third bucket is the one that matters
It is also the one the first version of this article got wrong. A citation to an agency's blog post about other agencies is not the assistant recommending that agency. Counting it as one is how you get a number like 88.
A quarter of what they read is a company
All 172 citations, by what the cited page actually was
| What the citation pointed at | Citations | Share |
|---|---|---|
| Directory or marketplace | 64 | 37% |
| A list on some company's site | 47 | 27% |
| A company's own website | 45 | 26% |
| Social, forum or news | 9 | 5% |
| A SaaS product | 7 | 4% |
Those 45 citations resolve to 44 distinct companies. Per question it is smaller still: 5 companies on the AI-first question, 13 on booking systems, 9 on European AI assistants, 7 on Lisbon, 10 on Portugal.
And 44 is the generous reading. It counts Capgemini, which is a global consultancy rather than anyone's idea of a small-business agency. It counts Peerbits, Concetto Labs and Tapadia Tech, Indian firms that rank for Portugal because they publish a Portugal landing page. It counts one individual freelancer's city page. Strip those out and 38 companies in the whole study were put forward as somebody who could do the work.
Three quarters of what these assistants read to answer "who should I hire" was written by somebody other than the people who would do the job.
The four are not doing the same job
The averages hide the finding. Split the 172 citations by assistant and the four of them behave like different products answering different questions.
| Assistant | Sources cited | Pointed at a company | Share |
|---|---|---|---|
| ChatGPT | 25 | 19 | 76% |
| Gemini | 46 | 14 | 30% |
| Claude | 23 | 6 | 26% |
| Perplexity | 78 | 6 | 8% |
Perplexity cited more than three times as many sources as ChatGPT and found six companies while doing it. It reads widely and almost entirely sideways, into directories and other people's rankings. ChatGPT cited a third as many sources and three quarters of them were companies.
Claude is a third case. It cited the fewest sources of anyone, and on the European AI-assistant question it cited nothing at all and answered from what it already knew.
This changes the practical advice. "Ask two assistants" is not useful on its own, because you need to know which two. If you want companies to contact, ChatGPT in this run was the only one reliably handing them over. If you want to understand the shape of a market, Perplexity's breadth is the point, and the output is a reading list rather than a shortlist.
Name a place and you get a directory
One pattern in the data is clean enough to act on. How much of an answer comes from directories and marketplaces tracks almost exactly with whether the question names a place.
| Question | Names a place? | From directories |
|---|---|---|
| Best AI-first web development agencies | No | 13% |
| Agencies building booking systems in the EU | Region only | 19% |
| Best agencies in Europe for AI assistants | Region only | 28% |
| A developer in Lisbon | City | 71% |
| Small software studios in Portugal | Country | 63% |
Ask where to find somebody and the assistants go to Clutch, The Manifest, GoodFirms, Twine, TechBehemoths, Sortlist and Upwork. Ask who is best at something and they go to other agencies' blog posts.
One site can own a question. On the European AI-assistant question, aiagencies.eu supplied 9 of the 36 citations, seven of them to Perplexity alone. A quarter of everything read on that question came from one directory.
If you are buying, a place-based recommendation is a directory lookup performed on your behalf. You can do that lookup yourself in less time and see the whole list rather than three names off the top of it. Worth knowing before you treat the three names as a verdict: several of the directory pages cited were a year old, including Clutch and GoodFirms rankings dated 2025.
If you are an agency, a directory profile buys visibility on "find me someone in Lisbon" and close to nothing on "who is best at AI integration", because on those questions the assistants were barely reading directories at all.
Who actually wins: the agencies that write about agencies
47 of the 172 citations were a list published on some company's own site. On the AI-first question it was 28 of 40. Claude cited seven sources on that question and every one was somebody's listicle. Gemini cited seven and six of them were, the seventh being a forum post.
So the companies getting read on capability questions are not the best agencies. They are the agencies with a content team.
Two examples make this concrete, and they are the two we originally and wrongly reported as the agencies the assistants agreed on. Musemind was cited by Gemini and Perplexity. Reverbico was cited by ChatGPT, Perplexity and Claude. In every case the citation was to a /blog/ URL where that company ranks other companies. Neither company's own website was cited by anyone, on any question.
They are not agencies two assistants agreed on. They are publishers two assistants read. Both are real companies with no connection to us, and they are named here because the pattern is not visible without them.
On a capability question, a decent share of the companies you are shown were selected by a competitor's marketing team, inside an article that competitor wrote to rank for the same question you just asked.
The question none of them understood
Question 2 asked which agencies build custom booking systems for small businesses in the EU. The answers came back full of booking products: Timify, Zeeg, Alteg, Carimeet, Glide. The question asked for a builder and the assistants answered with things to rent.
Glide shows the mechanism. Perplexity cited five separate Glide pages on that one question: the booking page for entrepreneurs, for agencies, for creative agencies, for digital marketing agencies, and for small business owners. Glide did not win that question by being the right answer. It won by publishing a page per audience segment, so whichever way the question was framed there was a page sitting on it.
This is what a category gap looks like. The companies exist, and the question has obvious commercial intent behind it. But nothing currently connects that phrase to a set of builders, so the assistants substitute the nearest category they do have, which is booking software.
For what it is worth, we are one of those builders, and we publish what a custom booking system actually costs.
Where we came out
BZB Studio was cited zero times in 20 answers.
The domain was four days old when this ran, with no directory listings and almost nothing linking to it, so the zero says more about the site's age than about the market. We are publishing it because an agency that measures a market and finds itself at the top of it has told you something about its method rather than about the market.
It also sets a dated baseline. We will run the same five questions again once the listings and the writing have had time to be crawled, and publish the second number next to this one whichever way it moves.
What to do with this
Pick the assistant for the job
ChatGPT handed over companies, Perplexity handed over reading lists. If you want names to contact, that difference matters more than asking twice.
Ask where the answer came from
Most assistants will show you. Three quarters of the sources here were written by somebody who is not the company being recommended, and a good share of those by a direct competitor.
For a location search, skip the middleman
Go to Clutch or The Manifest directly. That is what the assistant did, and you get the full list instead of three names off the top of it. Check the date on the page while you are there.
If you are an agency, list where location questions are answered
The Manifest was the most-cited single domain at 11, then aiagencies.eu at 9 and Clutch at 7. The vertical directory for your category can outrank the general ones.
Capability questions are won by publishing
On the AI-first question, 28 of 40 citations were somebody's blog post ranking agencies. The page that gets cited answers the question. It is not your home page.
Watch for the phrases with no companies attached
Question 2 had real commercial intent behind it and no set of builders attached to it, so a SaaS product with a page per audience segment took the whole question. Those gaps close eventually, and whoever is the clearest answer when they do gets the citations.
Limits and raw data
The raw export is published in full: ai-agency-citations-2026-09-25.csv, 172 rows, one per cited source, with the question, the assistant, the page title, the URL and the bucket we put it in. Every number above can be recounted from it, including the classification, which you are welcome to disagree with row by row.
What this study cannot tell you
It measures citations, not recommendations. An assistant can name a company without citing it, and that name does not appear in our data. This is the binding limit on everything above, and it is why the piece talks about what assistants read rather than who they recommend. One run per question, so sensitivity to phrasing and plain run-to-run variation cannot be told apart. One collection window, roughly two minutes on 25 September 2026, 19:08 to 19:09 UTC. Models change without notice and retrieval changes daily. Five questions, enough to show a pattern across question types, not enough to characterise any one market. We build the tool, which is why the export is published rather than described.
Method, stated plainly: five questions, four assistants, one answer each, United States market setting, 25 September 2026, collected with Allmond. The US market setting is a limitation worth naming, since four of the five questions ask about Europe; a European setting would likely raise the directory share further on the location questions. Geography was written into the question text rather than set as a regional preference, which is how a real buyer types it. Model versions and retrieval settings were not recorded, which we will fix on the rerun.
Questions people ask
When an AI assistant recommends an agency, where does that come from?
Mostly from somebody else's list. Across 172 sources cited in answer to five agency-buying questions, 64 were directories, 47 were lists published on some company's own site, and 45 pointed at a company's own website. On questions that name a city or country, directories supplied between 63 and 71 percent of everything cited.
Do all the AI assistants work the same way?
No, and the gap is wide. ChatGPT cited 25 sources and 19 of them were a company. Perplexity cited 78 and 6 of them were. Claude cited 23, and on one question cited nothing and answered from memory. If you want names to contact, the assistant you pick matters more than how many you ask.
Does being listed on Clutch help an agency appear in AI answers?
On location questions, yes. When a question named a city or country, directories supplied 63 to 71 percent of the citations, and The Manifest was the most-cited domain in the whole set. On capability questions the directory share fell to between 13 and 28 percent, and the citations went to other agencies' blog posts instead.
Can I trust an AI assistant's agency recommendation?
Treat it as one sample. Ask where it came from. On a location question you are reading a directory at one remove, and you can read the directory yourself. On a capability question you may be reading a list a competitor wrote.
Find out what the assistants say about you
Tell us your brand, your website and your industry. We run the questions your customers actually ask across all four assistants and send back what comes out, including the parts you will not like.
