“AI Search Is In The Plan.” Right, And Then What?
Somebody added a chapter called AI search to next year’s plan. Plenty of companies have got that far, and the heading looks respectable enough on the page.
Read that chapter to the end, though, and there’s often no line saying how anyone would know it’s working.
I wrote a chapter like that last year. The tactics ran to half a page; in the box for measurement I typed “we’ll work it out as we go” (which seemed like a perfectly sensible thing to type at the time), and then six months went by with the box still empty.
So maybe that’s a personal failing of mine. But somebody put the question to more than 600 marketing and public relations people, and it turns out my empty box is close to standard.
How Many Companies Actually Check Whether Their AI Search Work Is Doing Anything?
45% say they verify how visible they are in AI answers. On the sources an assistant cites, 60% say they don’t analyze them at all.
And set those two against the plans: according to Scrunch, whose survey the figures come from, 91% of the same group have an AI search strategy written down.
Nearly everyone can produce the document. Just under half as many can produce a number.
The survey was published on 16 July 2026, and it’s self-reported all the way through — more than 600 marketing and public relations practitioners in the United States (in-house teams and agencies mixed together), answering questions about their own work.
Back up to that 60% for a moment. “Citation sources” are the sites an assistant lists as the basis for what it just told you: the bibliography at the bottom of the answer. Which, for you, means the difference between your own page feeding an answer about your category and a competitor’s review page feeding it — and I’d go off and do very different work depending on which of those turned up.
“We Can Move This” And “I’m Not Sure We’re Right”, In The Same Building
84% of the same respondents think they can influence what the models say about them. 51% aren’t confident the strategy they’ve got is the right one. The competition question points the same unhappy way: 62% feel they’re behind, which makes for an odd set of beliefs to be holding at once — we can move this, we don’t know whether we’re moving it correctly, and everyone else is further along.
The obvious move from there is to buy something, and 73% report they’ve already invested in tools, against the 45% who verify. Since both figures come from the same 600-plus people, the same group is telling me it owns something it isn’t opening.
So why does the tooling number sit above the checking number?
Because buying and reading are two different acts, and only one of them goes on a purchase order.
A subscription buys access, and access is where a lot of us stop; the direct debit for a newspaper goes out whether or not anybody ever opens it.
What Do You Count To Check Whether You’re Visible In AI Answers?
“Verify” means whatever the person saying it has decided it means, so the definition is worth pinning down before anyone goes shopping for a tool that supplies one. Three things are on my list, each with a cadence attached to it:
- Mention rate — put the same question to a model over and over, then count the answers your name turns up in. Monthly.
- The citation line-up — write down which domains the answer cites, and what share of them are trade press, your own site, or posts by users. Monthly.
- Spread between models — the same question to ChatGPT, Claude, Gemini and Perplexity, then read where they disagree. Quarterly.
Mention rate is the one worth spelling out, since the phrase does no work on its own: how many times, out of how many asks, your name comes back. For example, twelve mentions across fifty asks is 24%, and 24% is a thing that can go up next month.
Or take it the other way round. Ask once, and the only thing on the table is the running order inside a single answer — and one answer tells you about as much about your standing as one coin toss tells you about the coin (I have put exactly that single coin toss on a slide, with an arrow pointing at it).
When any of the three goes into a deck, I keep the conditions in the same sentence as the value: how many asks, which model, which date. That way next quarter’s version is comparable with this one, and comparability is the entire reason for counting anything.
How Far Does This 45% Travel?
Every figure above came out of a form people filled in about themselves. Nobody read a log of the questions actually put to a model, and nobody counted company names in the answers that came back — which would be a different kind of evidence altogether, and a much harder one to argue with.
The people filling in that form were in the United States: 602 of them. What the equivalent split looks like in your own market, this survey doesn’t say, and I wouldn’t assume it transfers.
Method is partly on the record. Fieldwork ran from 19 May to 2 June 2026, and the results are reported at 95% confidence with a margin of error of ±4% (the width of the wobble to expect around each percentage). But that ±4% belongs to proportions drawn from a US sample, so it travels exactly as far as the sample does and not a step further.
What isn’t on the record is how those 602 people were approached in the first place. If a sample’s origin is invisible, its lean is invisible too: a panel recruited from buyers of AI search tooling and a cold list pulled off a professional network would never hand anyone the same 45%.
Then there’s who published it. Scrunch sells AI visibility measurement, so “not enough people are measuring” is a conclusion that happens to describe the market for its own product. But I don’t think that makes the figures wrong. I’d just like to see somebody with nothing to sell run the thing again (and I’d want the same if a study of ours ever landed this conveniently).
One more gap nags at me. The survey doesn’t define what “verifying” or “analyzing citation sources” were taken to mean, and without a definition the person who glances at a dashboard once a quarter and the person running a monthly count both tick the same box. So a fair slice of that 45% could be doing the lighter of those two, and nothing in the survey rules that out.
What survives every one of those is the shape. In one sample, answering one set of questions, holding a strategy and checking it can sit about two to one apart — and a shape is something I can hold up against a plan today, without waiting for anyone to run the study again.
Add The Counting Line Before You Add Another Tactic
One line does it.
What gets counted, how often, on which model — written into the plan alongside the tactics, in the same size type, with somebody’s name against it. I’d pick whichever of the three above would change a decision I actually make. Mention rate and the citation line-up at a monthly cadence put six points on a chart by the time the half-year review comes round, and six points is enough to argue with.
Maybe I’m making too much of a form people filled in about themselves. Still, the part I’d defend is narrower than the survey: in a market where nearly everyone has the strategy written down, what’s lying around unclaimed is the evidence that any of it works — and evidence is what turns AI visibility from a chapter heading into something a brand can improve on purpose.
Whether the ratio holds at 91 to 45 outside the United States, I’ve no idea. Whether your own plan carries that line, you can settle this afternoon.
Sources
- Eric Wendt (Scrunch), “The AI confidence trap: Key takeaways from our 2026 AI search survey”, Scrunch Blog, 16 July 2026 (limitation: Scrunch sells AI search visibility measurement tooling, so it holds a commercial interest in the conclusion that measurement is thin on the ground)
- Scrunch and Scribewise, “Moving fast, flying blind: What 600-plus marketing and PR professionals reveal about the gap between AI search confidence and AI search competence” (the 2026 AI Search Survey itself), Scrunch, last updated 16 July 2026 (self-reported survey of 602 marketing and public relations practitioners in the United States, fielded 19 May to 2 June 2026, reported at 95% confidence with a margin of error of ±4%; how respondents were recruited is not published)
