Fundamentals

AI named the same firm the best MSP in eight states. It is a staffing agency.

By Arnav Mukherjee, founder of TofuBofu · August 14, 2026

The most frequently recommended managed IT provider in North America, by the measure of our own research, is a recruiting firm in Philadelphia that does not run anyone's IT.

On 12 and 13 July 2026 we put one open question to each of the six AI engines we track, in each of fifteen North American regions: who are the best managed IT service providers here. Four of the six answered. We took every company name the engines volunteered, 566 firm-and-region entries in all, and resolved each one to a real website. I went back into that data this week for an unrelated reason and stopped on the name that recurs most often across the fifteen regions. It appears in eight of them. It is Symmetrio, and its own homepage describes it as "a full-service recruiting, staffing, and consulting company" serving the Philadelphia region.

Then I stopped treating it as an anecdote and checked every other name in one of those answers. That is where the useful part is, and it is worse than one odd recommendation.

Twelve names, checked one by one

Here is Claude's stored answer for California, verbatim, because the full list is far more instructive than the single name that caught my eye.

I opened each of those twelve companies' own websites this month and read how they describe themselves. Every claim below is that company's own words about its own business, and none of it is a criticism of the company. These are real firms doing real work. The engine put them on a list they never asked to be on.

Start with the middle group, the one labelled as the California-based specialists a local buyer would actually shortlist. Webris is a marketing agency: "we partner with law firms to build marketing campaigns that generate signed cases". Splashtop sells "top rated remote access, support, and endpoint management software". Fortive describes itself as a global team that solves "critical challenges in industrial and healthcare settings". Symmetrio is the Philadelphia recruiter. And crewlogic.com, the domain our resolution pass matched to the fifth name, now redirects to a marketplace listing offering the domain for sale.

Five names offered as California managed service providers. Not one of them is a California managed service provider.

The first group is a different failure. Kaseya's own site says its platform "transforms how MSPs and businesses operate, secure, and grow" and calls its products "powerful solutions built for MSPs and IT teams". It sells software to managed service providers. Datto, which the answer itself correctly notes was acquired by Kaseya, tells IT professionals its solutions "drive cyber resilience, efficiency, and growth for MSPs". Recommending those two as the best MSPs in California is like answering a question about the best restaurants in town by naming the food distributor.

Of the whole twelve, two survive contact with their own homepages. Complete IT is a genuine managed IT provider and lists offices in Camarillo, Irvine and Costa Mesa, so it belongs on this list. CDW is a real national technology provider with California operations, which the answer flags itself. The rest include a healthcare software company (nextgen.com is NextGen Healthcare, selling electronic health records to medical practices), a name our resolution pass could only match to a computer-training listing on the other side of the world, and the seven above. A buyer reading that answer and calling three names off it has a fair chance of reaching a recruiter, a law-firm marketing agency and a domain that is for sale.

A different head office nearly every time

All eight Symmetrio placements came from one engine, Claude, and the detail that gives the game away is the location attached to each. Asked for the best providers in California it offered "Symmetrio (Bay Area)". In Florida, "Miami-based, focuses on managed services and cloud". Then "Chicago-area IT services" in Illinois, "New York operations", "Columbus-based managed services" in Ohio, "Dallas-based" in Texas, and in Pennsylvania, the one state where the company actually operates, "Symmetrio (Pittsburgh)", which is the wrong end of it. In New Jersey the answer skipped the geography and described it as "focused on mid-market businesses".

Seven cities for one firm. What the engine produced there was not a recalled fact with an error in it. It was a name it had learned to associate with IT work, fitted to whatever geography the question supplied, with a plausible city generated to finish the sentence.

The same engine does it again with a second Pennsylvania company. Compunetix describes itself as "the leading developer of converged VoIP, voice, video, and data collaboration and conferencing applications", and it is headquartered in Monroeville, Pennsylvania. Claude named it as a managed IT provider in four regions: "Edmonton-based" in Alberta, "Vancouver-based, serves enterprise clients" in British Columbia, "Based in Winnipeg, offers managed services" in Manitoba, and "Operating in North Carolina". Three Canadian head offices, in three provinces, for a conferencing hardware developer outside Pittsburgh.

That second case is what turns the first from an anecdote into a pattern. The engine is not confused about one company. It is doing a consistent thing: producing a name that sits near the category, then decorating it with the local detail the question invites.

Category adjacency is not category membership

This is our read rather than a measurement, and it is the reusable half of the story. An AI engine has no register of who sells what. It has text. When it assembles a list of the best providers in a category, it reaches for names that appear in the linguistic neighbourhood of that category, and a staffing firm that places IT engineers lives in exactly that neighbourhood. Its pages are dense with infrastructure, engineers, support, cloud, the names of client industries. Every signal a model can read says IT. Nothing on the site says "we do not run your IT, we hire the people who do", because why would anyone write that down.

Category membership, for almost every B2B company, is implied by context and stated explicitly by nobody. A human reader infers it in about a second from a homepage: the photography, the pricing page, the shape of the case studies. A model working from text has none of that and does not get the second. So it draws the boundary itself, and it draws it around vocabulary.

The pattern repeats across the study, not just inside one answer. Kforce turns up as a recommended provider in two regions, and Kforce describes itself as "a solutions firm specializing in technology, finance and accounting, and professional staffing services", placing people rather than running infrastructure. Kaseya and Datto each appear in two. ConnectWise, another vendor selling into the category, appears in one. That is two staffing firms and four software companies collecting placements on lists of the best firms to hire for your IT.

One group I deliberately left out, because including it would have been the easy version of this argument. Telus Business, Bell Business and Rogers Business also recur across the Canadian regions, and it is tempting to file three phone carriers alongside the recruiters. That would be wrong. Telus bought Fully Managed, a real mid-market MSP, for roughly 137 million dollars in January 2022, and our own analysis script carries a hand-written alias mapping "Fully Managed by TELUS Business" onto TELUS Business. The carriers genuinely sell managed IT. They belong to a separate complaint about national incumbents crowding regional lists, which we have written about elsewhere, and mixing the two would have made a sharper-sounding post out of a weaker claim.

The strongest objection: this is one weak engine

Everything above came from Claude, and a fair reader will say so before I do. Claude had the weakest grounding of the four engines in this study by a distance. Of the names it produced across fifteen regions, 2 of 129 could be corroborated against the region's own directory or a live web search. Perplexity managed 101 of 257, Gemini 24 of 137, ChatGPT 8 of 114. Picking on the bottom of that table and calling it a finding about AI engines would be dishonest.

So I ran a check that could not be steered by which answer I happened to read first. Our company-resolution script carries a fixed list of forty well-known software vendors, written months earlier for an unrelated purpose and never touched since. I matched every name each engine produced in these fifteen regions against that list, on whole words. Claude returned 11 of its 129. ChatGPT returned 1 of 114. Gemini returned 0 of 137. Perplexity returned 0 of 257.

Perplexity was asked the same fifteen questions, produced twice as many names as anyone else, and did not once offer a software vendor as a managed service provider. That is the objection answered and conceded at the same time. The category-adjacency failure in this study is overwhelmingly one engine's, and the thing that separates that engine from the others is that it was not retrieving before it answered. An engine that reads live pages stays tethered to what those pages say about themselves. An engine reasoning from trained memory produces a fluent, locally detailed shortlist assembled from names that feel adjacent to the question.

The second obvious defence is to trust only names that several engines agree on. We tested that too. Of the 566 entries, 55 were named by two or more engines, and 29 percent of those cleared an independent check, against 19 percent of the entries a single engine produced. The direction is right. The size is not reassuring: seven in ten of the agreed-on names still failed. On a base of 55 that is suggestive rather than settled, and I would not build a buying process on it.

Which leaves the reader in a specific position rather than a vague one. If your buyers use Perplexity, published work moves your standing and the list they see is mostly composed of real firms. If they use an engine answering from memory, you are competing for space on a list that also contains recruiters and conferencing vendors, and no amount of being good at your job will dislodge them. Both facts point at the same job.

What this costs you, and the sentence that fixes it

If you run a genuine regional provider, the practical damage is not that a recruiter took your slot once. It is that the list your buyer is reading is shorter than it looks. Strip out the vendors selling into your industry, the staffing firms, the conglomerate and the domain that is for sale, and a twelve-name answer collapses to two real candidates. You are contesting far fewer places than the answer implies, which cuts both ways: the field is thinner than it appears, and so is your margin for being absent from it.

The composition of that list is also a diagnosis of your own content, and this is the part most founders skip. When the firms named beside you are not the firms you meet in competitive deals, an engine has framed your category from text, and your text was not among what set the boundary. It is the same signal as a competitor mismatch read from the other direction: instead of your content filing you in the wrong bucket, it has failed to stop other companies being filed into yours.

The fix is unglamorous and most sites do not do it. State the boundary in plain sentences, including the negative half. Not a services grid, not an icon row, a sentence a model can lift whole: what you are, who you serve, where, and what you do not do. "We are a managed service provider for firms of 20 to 200 people across Ohio. We do not resell hardware and we do not place contract staff." Schema markup lets a parser confirm that claim and is worth having, but the sentence is what gets quoted into an answer. Structure alone does not separate a cited firm from an uncited one, and we have published our own crawl data saying so.

Then check it the only way that means anything. Ask the buying question, on several engines, and read the full list instead of scanning it for your own name. Count how many of the companies on it are genuinely in your category and genuinely in your market. That count is the real measure of how contested your slot is, and it is a number no dashboard will hand you, because it takes a human opening twelve homepages, which is exactly what I did to write this.

Then fix the boundary, hold the question fixed, and ask again in a month. Changing the question between runs is how founders convince themselves nothing is moving. A shortlist you have never read all the way through is a shortlist you cannot argue with.