Duane Forrester Says You Won’t Be Able to Check Your 2027 GEO Metrics. Here’s What You Still Can Count.

Trend watch — published September 18, 2026. Duane Forrester (ex-Bing, ex-Yext) published “Three Predictions for 2027, and Why You Won’t Be Able to Check Them” on September 16 (Search Engine Journal, and on his own Substack). We pulled his exact framing, tested it against the AI-search metrics we run every week, and asked the practical question he leaves open: if the dashboards go dark, what can you still measure? — in the spirit of the GEO Lab.

What did Duane Forrester actually predict?

Three things, and they stack. First, that search advertising revenue keeps growing in absolute terms while the share of it earned by sending a person to somebody else’s website keeps shrinking — intent gets monetized where it’s expressed, inside the answer, and the ad business grows anyway. Second, that the ranked, ten-blue-links results page persists mostly as a legacy interface, kept alive by advertisers and two decades of habit, while the consequential product decisions move somewhere else. Third — the one that matters for measurement — that “the surface that ranks and the surface that earns are decoupling,” and most organizations spend 2027 running in parallel: optimizing a ranked surface and an answer surface at the same time, unsure which one is paying the bills.

His grounding fact is worth repeating because it’s counterintuitive: as Forrester notes, Google’s Q2 2026 “Search and other” advertising revenue came in around $63.3 billion, up roughly 17% year over year — a full year into AI Overviews and the global rollout of AI Mode. The zero-click answer layer didn’t dent the money. That’s the whole tension: the business is healthy, but the instruments you’d use to attribute your slice of it are quietly going dark.

What’s the difference between a counted number and an inferred number?

This is the sharpest line in the piece, and it’s a better mental model than the ones we’ve been using. In Forrester’s framing: “A counted number comes from an event that was observed on infrastructure you or your vendor control. An inferred number comes from a sample extrapolated to a population nobody can fully enumerate… Only one of them can be checked.” A checkout on your own site is counted. A “you appear in 34% of AI answers for your category” figure is inferred — it comes from someone sampling a set of prompts they chose and projecting the result onto every prompt that exists, a population no one can list.

Both kinds of number can be useful. The danger is reporting an inferred number with the confidence of a counted one — drawing a trend line, setting a KPI, moving budget — when the thing underneath is a sample you can’t inspect, produced by a method you can’t audit, projected onto a universe you can’t enumerate. Forrester’s test for any vendor number is three questions: What population does this describe, and can it be named? Was the value observed or extrapolated? What would move this number if the world outside stayed exactly the same? If the answer to the last one is “the vendor changed their sample or their model,” you’re looking at inferred.

Which GEO numbers are inferred — and can’t really be checked?

Most of the ones the industry is currently selling as scoreboards. We’ve hit this wall repeatedly in the lab, and Forrester’s label explains why each one felt slippery:

• Any “AI visibility %” or “AI share of voice” from a prompt sample. The vendor picks the prompts; you can’t enumerate all the ones real users type. It’s an extrapolation, which is exactly why a blended “AI Share of Voice” number is the wrong metric to run a strategy on.
• Vendor “AI traffic converts Nx better” multipliers. When we chased the popular “4x” stat, the reported range ran from −13% to 23x depending on who counted and how — a 300-to-1 spread is the signature of inferred numbers built on tiny, self-selected denominators.
• Search Console’s AI-surface “position.” Google itself just clarified that every link in an AI Overview gets the block’s position, not a rank — a block-level stand-in, not an observation of where you sit in the answer.
• Cross-engine “leaderboards.” Because engines barely overlap — our testing found the four major ones shared only 2 of 395 cited domains, with ~85% unique to a single engine — any single blended visibility rank is smoothing over four different populations, which is why there is no universal GEO strategy to score against.

None of these are useless. They’re directional — a compass, not a scoreboard. The error is precision: reporting an inferred visibility score to a decimal and defending it in a QBR as if it were a counted fact.

Which GEO numbers stay countable?

Two categories survive Forrester’s test, and they happen to be the two we lean on hardest. First, your own first-party events. A session that arrives from an AI engine and lands on your server, the page it reads, the form it fills, the order it places — those are observed on infrastructure you control. That’s why we keep pointing people to a proper GA4 setup for AI-referred traffic instead of a vendor dashboard: it’s counted, it’s yours, and it answers the question that actually pays — did the visit do anything? The one honest caveat is coverage, not method: some engines strip the referrer (ChatGPT’s app traffic often shows as “direct”), so first-party data undercounts AI’s contribution. But what it does capture, it captures as fact.

Second — and this is the part the vendors can’t sell you — your own controlled experiments. You can’t inspect a platform’s ranking method, but nothing stops you from running your own method, in the open, on a population you define. Change one variable, hold the rest still, observe the citation outcome across engines, publish the verdict. That’s not a workaround; it’s the entire GEO Lab. When we wanted to know whether being listed in an AI answer meant being picked, we didn’t buy a score — we ran the test and found listed is not the same as picked and later cited is not the same as recommended. Forrester’s warning is “methods you cannot inspect.” The answer is a method you can inspect: yours.

So is AI-search measurement hopeless?

No — but the center of gravity has to move. For twenty years the industry outsourced its scoreboard to the platforms: Search Console told you your rank, analytics told you your clicks, and you built reporting on borrowed numbers. Forrester’s prediction is that the borrowed numbers are becoming un-auditable faster than they’re becoming useful. The teams that will still be able to defend a decision in 2027 are the ones that already do two things: instrument their own events, and run their own experiments, treating every platform-supplied number as a directional input rather than a source of truth. That’s not pessimism about GEO. It’s just moving from “what does the dashboard say” to “what did I observe.”

The scope discipline is the same as always. This is a measurement argument, not a GEO tactic — nothing here changes what gets you cited (answer-first structure, first-party data, category authority still do that across engines). And because engines diverge, “measure it yourself” means per-engine, not one blended figure; the lightweight coverage passes in our per-engine GEO checklist are counted observations you control, one engine at a time.

The practitioner takeaway

1. Label every GEO number counted or inferred before you report it. First-party events are counted; anything sampled-and-projected is inferred.
2. Report inferred numbers as direction, not decision-grade. No decimals, no week-over-week trend lines, no budget moves on a visibility score alone.
3. Ask Forrester’s three questions of any vendor stat. What population, and can you name it? Observed or extrapolated? What would move it if the outside world stayed still?
4. Instrument your own events. A GA4 setup for AI traffic gives you counted sessions and conversions — undercounted by referrer loss, but real.
5. When a number matters, run the experiment yourself. One variable, cross-engine, published verdict. The method you can inspect is the only one you fully control.

Bottom line: Forrester isn’t predicting the end of measurement — he’s predicting the end of borrowed measurement, and he’s mostly right. The dashboards will keep producing confident-looking numbers built on samples you can’t inspect and populations you can’t name. The move isn’t to stop using them; it’s to demote them to what they are (a compass) and to build your actual decisions on the two things that stay countable: the events that hit your own infrastructure, and the experiments you run in the open. If your reporting already separates “we observed this” from “someone estimated this,” 2027 doesn’t change much for you. If it doesn’t, this is the year to start labeling.

Frequently asked questions

Is Duane Forrester saying AI-search analytics are useless?

No. His argument is that by 2027 more decisions will rest on inferred data — samples extrapolated to populations nobody can fully enumerate — reported with the same confidence as counted facts. The takeaway isn’t to abandon AI-search analytics but to treat inferred numbers as directional inputs rather than decision-grade truth, and to base consequential calls on numbers you can actually check.

What is the difference between a counted and an inferred number?

A counted number comes from an event observed on infrastructure you or your vendor control — a session on your site, a conversion, an order. An inferred number comes from a sample extrapolated to a population nobody can fully list, such as “you appear in 34% of AI answers for your category,” which projects a chosen set of prompts onto every possible prompt. Only the counted number can be independently checked.

Can I measure AI search visibility at all, then?

Yes — with two categories of number that stay countable. Your first-party analytics (AI-referred sessions, conversions, and revenue that reach your own servers) are observed events you control, though referrer loss means they undercount AI’s true contribution. And your own controlled experiments — change one variable, observe the citation outcome across engines, record the result — give you a method you can inspect, unlike a vendor’s black-box score.

Is Google’s Search Console AI report a counted or inferred number?

Mixed. Its impressions count real events on Google’s surfaces, but it has no click data and covers only Google, so it is counted-but-partial. Its AI-surface “position,” by contrast, is a block-level stand-in — every link in an AI Overview shares the block’s position rather than an observed rank — so it behaves more like an inferred number and should not be charted as a ranking.

Should I stop using AI visibility tools?

No — use them as a compass, not a scoreboard. Third-party AI visibility scores are directional signals about where you might be showing up, which is genuinely useful for prioritization. Just don’t report them to a decimal, build KPIs on them, or shift budget on them alone. Pair them with first-party event data and your own experiments so the decisions rest on numbers you can check.