Google Trends is probably not "broken," but in 2026 it deserves more scepticism than it used to. Several practitioners have flagged mature, stable categories — air fryers, car insurance, London hotels — showing all-time-high "interest" that doesn't square with real-world demand. One widely-shared hypothesis is that AI query fan-out is inflating the signal by counting machine-generated sub-queries as human searches. Google has not confirmed any cause, so the honest read is: treat Trends as directional, not ground truth.
What's actually being claimed?
The observation comes from practitioners, not from Google. The pattern they describe: keyword tools and Google Trends show record interest in well-understood, slow-moving categories where demand should be flat. A category like air fryers or car insurance has no reason to hit an all-time high — the products and the buyers haven't changed.
One practitioner's hypothesis, shared in a widely-circulated post, ties this to how AI search works. When an AI assistant answers a question, it often doesn't issue one search — it fans the question out into many backend sub-queries against Google and other sources. If those automated sub-queries land in the same measurement pipeline as human searches, the volume for terms no human typed gets inflated.
This is a hypothesis. It is plausible and it fits the pattern, but Google has not endorsed it, and other explanations (sampling changes, bot traffic, methodology shifts) are possible. State it as a hypothesis when you repeat it.
How could query fan-out inflate the numbers?
Query fan-out is the mechanism behind modern AI answers: one question becomes many searches.
- A user asks one question in an AI surface.
- The system decomposes it into sub-queries — rewrites, comparisons, anticipated follow-ups, slot-filling for variables like price or location.
- It searches each sub-query, often in parallel, and synthesises the results into a single answer the user reads without clicking.
Independent analyses estimate that Google's AI Mode can fire on the order of ten sub-queries for a single question, while chat engines like ChatGPT tend to issue fewer — figures that vary by query and are not officially published, so treat them as rough estimates. The point holds either way: one human question can generate many machine searches. If a measurement system can't cleanly separate the two, aggregate volume and "interest" rise without any change in human demand.
Are humans really searching less?
The distortion hypothesis lands harder because human search behaviour appears to be moving the opposite way. According to clickstream analysis from Datos and SparkToro (reported mid-2026), Google searches per US user fell nearly 20% year over year — attributed largely to AI answers and instant results satisfying intent without follow-up searches. The same analysis put the decline in Europe at only about 2–3%, a much smaller shift.
If human searches per person are falling while headline "interest" hits record highs, the gap is the tell: the curve may be tracking machine activity, not people. Read it as a signal to corroborate, not a number to plan against.
Treat both figures as vendor-reported and directional — but together they sketch a coherent picture: fewer human searches, more machine sub-queries, and aggregate signals that no longer map cleanly to real demand.
What does this mean for brands?
Two kinds of decisions lean directly on keyword and Trends data, and both are exposed:
| Decision | Why it's exposed | Safer practice |
|---|---|---|
| Content planning | Topic priorities chosen by search volume may chase inflated terms | Plan around the questions buyers actually ask AI, not raw volume |
| Market sizing | Demand estimates built on keyword volume may overstate a flat category | Triangulate with first-party demand signals and sales data |
| Trend spotting | An "emerging" spike may be fan-out noise, not a real shift | Confirm against a second, independent source before acting |
The takeaway is not to abandon keyword data — it's still useful for the shape of demand. It's to stop treating it as the single source of truth in a zero-click, conversational-search world where much of the measured activity may be machines, not buyers.
What should you trust instead?
Pair traditional keyword and Trends data with answer-level signals — the part of discovery that now happens inside AI answers. The questions worth tracking:
- Citation share — across a representative set of prompts your buyers would ask, how often does an AI engine name or cite you?
- Recommendation rate — when the question is "best X for Y," are you in the set the engine recommends?
- Movement over time — is your share rising or slipping across engines, sampled repeatedly?
These signals measure whether real demand is reaching you, independent of whatever the search-volume pipeline is counting. (For the broader migration from search clicks to AI citations, see from clicks to citations; for the underlying numbers, the AI search statistics 2026 reference keeps every figure attributed and dated.)
When keyword tools may be counting machine traffic, the more direct read on demand is whether AI engines actually surface and recommend you — sampled across many prompts, repeatedly, on every engine. That measured answer-level visibility is exactly what Buffy Intel is built to track.
