Signals

We crawled 14 firms AI names and 14 it ignores. The checklist barely told them apart.

By Arnav Mukherjee, founder of TofuBofu · August 11, 2026

The most common question I get from a founder holding a bad scan is some version of this: what does the other firm have that I do not? Every AEO guide in the category answers it the same way. Add Organization schema. Add LocalBusiness. Ship an llms.txt. Mark up your FAQs. Build out city pages. The list is always the same list, and it is always presented as the difference between being recommended and being invisible.

So we tested it against our own data. For each managed IT firm our scans had already measured as absent from AI vendor answers, we took the specific rival the engines named instead, in the same region, in the same category, and crawled both sites the same way on the same day. Not a firm we assumed was doing well. The firm the engine actually said the buyer should call.

Fourteen usable pairs came back. The checklist did not separate them. Twelve of the fourteen invisible firms already carried Organization markup. More of them served an llms.txt than the named firms did. More of them published pricing pages. On the axis every guide leads with, the firms nobody recommends were, if anything, slightly ahead.

How the pairs were built

The starting list came from scans we had already run: firms where the engines answered a regional buying question and never said the firm's name. For each one we took a rival that had been named in that firm's own answers, resolved it to its real website, and crawled both. Same crawler, same depth, same day.

We started with 29 pairs and kept 14. The eleven we dropped are worth a sentence, because they are their own finding. In every dropped pair one side returned almost nothing: a hard block, a timeout, or a rival that no longer exists independently. Two of the named rivals had been acquired, and their domains now redirect to the buyer's site. The engines are still recommending them by name.

Three more things you need in order to read the table honestly. The 14 invisible firms are 14 distinct companies, but the named side is only 8 distinct companies, because several of them get recommended repeatedly across a region. Our crawler reads up to 20 pages and 200 sitemap URLs and caps each page-type count at six, so every number below is a floor, not a total. And these firms were selected because they were already invisible, so this describes the invisible group, not the industry.

What the two groups actually looked like

Counted once per firm, so the repeated rivals cannot inflate anything, here is the share of each group carrying each item.

Read the top of that table slowly, because it is the part that costs people money. Organization schema is the first item on almost every AEO checklist published in the last two years, and it is present on 86 percent of the firms no engine will name. It is not a differentiator. It is table stakes that both groups already cleared.

Three rows run backwards. More invisible firms serve an llms.txt than named ones. More of them publish pricing. More of them have industry pages. I am not claiming those things hurt you, and a sample this size cannot support that claim. What it does support is narrower and still useful: if your plan for getting into AI answers is to work down that list, you are about to spend a quarter acquiring properties that the firms already beating you mostly do not have.

Three pairs that make the point better than the averages do

Averages hide the cases that should change your mind. These three did mine.

The one axis that moved, and how far it actually goes

Exactly one thing moved consistently in the right direction: questions published and marked up as questions. Five of the 14 invisible firms carried any FAQ markup, against five of the eight named firms. Among the sites that had it at all, the named firms had marked up roughly twice as many questions each, about 13 against about 7.

That is a thin result on a small sample and I am not going to dress it up. Nine of the 14 invisible firms had zero marked-up questions anywhere we could see, so most of what you are looking at is an absence on one side rather than a mountain on the other. But it points the same way as everything else we have measured. An engine assembling a shortlist is looking for text that already answers the question it was asked. A question written out, answered plainly, and labelled as a question is the most directly liftable form that text can take.

Notice what that is not. It is not markup as a technical property. The FAQ schema is the wrapper. The thing doing the work is that somebody sat down and wrote the answer to what a buyer asks. Our own crawl found the ceiling on this quickly: one firm has two FAQ pages and three case studies already written and published, carrying no markup at all and never saying who published them. They did the hard half and skipped the easy half.

The claim I am not going to make

There is a tempting story in this data and it does not survive contact with the sample. At pair level, the named rivals carry 54 city and industry pages against 30, and lead on that axis in five pairs against three. Read that on its own and you would publish a piece arguing that a page per question is the whole game.

Count each firm once and it falls apart. Only eight distinct companies fill those fourteen named slots, and two of them are unusually page-heavy, so the totals are mostly those two counted repeatedly. Per firm, half the invisible firms have industry pages against 38 percent of the named ones. The axis runs backwards the moment you stop double-counting.

I am including that because it is exactly the mistake I would have made reading the first summary of this crawl, and because it is the mistake most vendor research in this category is built on. A sample of eight distinct firms in one profession cannot settle anything. It can retire a bad assumption, which is what the top of the table does, and it can point at where to look next, which is what the FAQ row does. That is all it can do, and pretending otherwise would make the rest of this untrustworthy.

So what should you actually do

If your site already carries Organization markup and an llms.txt and you are still not being named, the checklist has nothing left to tell you. Four things do.

The framing we keep returning to is that search rank is the floor and AI visibility is a distinct layer built on top of it. This crawl is the cleanest evidence I have that the two are genuinely separate. Both groups had real websites, valid markup and working sitemaps. They cleared the floor. What separated them, to the small extent anything did, was whether the buyer's question had ever been answered somewhere an engine could retrieve it.

That matters more every quarter. G2's 2026 research found 51 percent of B2B buyers now begin vendor research on an AI chatbot, up from 29 percent, and Forrester's 2026 study found 94 percent use AI somewhere in the buying process. SE Ranking's 2026 analysis found 71 percent of ChatGPT-cited pages carry structured data, which is the sort of number that gets quoted as a reason to go and add markup. The firms in our sample had done that. It was not enough on its own, and knowing that is worth more than another quarter spent on the checklist.