The July Search Break
Eggs Benedict Tech · Field note 23 September 2026

Search measurement

The July search break

In one week, Google Trends stopped seeing half of AI search demand. Ordinary queries were untouched. Here is how to tell the difference — and what it costs anyone measuring this market.

What we found

Across 20 AI-related search terms, web search interest fell by a median of 50.9% in the single week of 12 July 2026 and stayed down. The same terms held on YouTube and rose on News. Four ordinary commercial queries on the same surface — car insurance, quickbooks, divorce lawyer, plumber near me — did not move at all. The break is specific to AI-intent queries on web search. Whether those queries stopped being counted or stopped being typed there, we cannot tell from outside, and both point in the same direction.

01 · The observation

A step, not a slide

Median one-week fall−50.9%Across 20 AI-related terms with a measurable level, week of 12 July 2026
Terms falling >25%16 / 20In that one week, across unrelated vendors, categories and phrasings
Same terms on video89.8%Share of pre-break level retained by “openai” on YouTube search

The fall arrives in a single week and then flattens. “ai automation services” went 56 → 18. “self-hosted ai” went 64 → 23. “openai” went 57 → 32. Post-break slopes are close to zero: this is a step down to a new plateau, not the start of a decline.

Searches for “openai”, web against video

Weekly · United States · Sept 2025 – Sept 2026 · each series indexed to its own peak

Web Search YouTube Search
One term, one country, one set of weeks, two surfaces of the same measurement system. A fault in the system itself would show up in both lines.

Three AI vendor terms, both surfaces

Weekly · each panel indexed to its own peak · dashed rule marks 12 July

Two of the three break on web and hold on video. “google gemini” breaks on neither — a useful complication we do not explain: its web series is flat across the period, which is what a term with a large and rising non-AI-intent base would look like.

02 · The tests

Four explanations, eliminated

A single break in a single dataset invites over-reading. Each of these tests was run to kill a specific explanation, and each of them killed one.

01

Seasonality

June to September in prior years, same term: 1, 1, 1, 1 in 2022 · 3, 3, 3, 3 in 2023 · 4, 4, 5, 8 in 2024 · 25, 33, 28, 33 in 2025. In 2026: 100, 71, 31, 19. No prior year declines into September.

02

A reporting-threshold artifact

Small terms do round to zero. But terms at 5–20, 20–50 and above 50 retained 21.5%, 14.3% and 15.2% — indistinguishable. Correlation between a term's pre-break size and how much it kept was only +0.37 across 21 terms. Size is not what decided this.

03

A fault across the measurement system

The same three vendor terms retained 83–90% on YouTube search and rose on News search over the same weeks. Two of three surfaces are unaffected.

04

A change to web search counting generally

Four non-AI controls on the same surface, over the same weeks, show no step at all. Whatever changed did not change web search measurement as a whole.

What survived the week of 12 July

Share of pre-break level retained · web search · mean of 10 May–5 July against 2 Aug onward

AI-intent query Non-AI control
The control terms were chosen to span intent types — an insurance shopper, a software brand, a professional-services need, a local trade — so that no single query style carries the comparison.
The property test · share of pre-break level retained
TermWeb SearchYouTubeNews
openai18.9%89.8%rose
anthropic42.1%90.1%rose
google gemini94.5%82.7%rose

03 · The reading

What is left, stated carefully

The break is specific to AI-intent queries, on web search, in one week. Two explanations survive, and this data cannot separate them.

People stopped typing those phrases into that box. Someone who would have searched “private ai assistant” now asks an assistant instead, in a sentence rather than a keyword. Conversational AI search, which lives on this surface, is the obvious candidate — reporting on it describes queries roughly three times longer than traditional ones, because people narrate context rather than typing keywords.

Or the queries still happen and are no longer counted the same way. A platform that answers a question in place may account for that interaction differently from a classic search, and only the platform can say. This class of change has visible precedent: when a search parameter was quietly retired in September 2025, third-party impression data fell off a cliff while nothing about demand had changed.

For anyone measuring this market from outside, the distinction matters less than it appears. Under either reading, the public instrument no longer registers demand it used to register. The series before July and the series after are not the same measurement.

There is a second-order effect worth naming, because it is easy to miss while looking at charts. Aggregate search interest has been, for two decades, an unusually democratic dataset: a rough public record of what people wanted, available free to anyone with a browser. When those questions move into a conversation with an assistant, that record thins. The demand does not disappear — it becomes private, held by whoever operates the assistant. That is better for the individual asking, and worse for everyone trying to understand a market from the outside, including the regulators, researchers and small businesses who never had access to anything else.

04 · Consequences

If you plan against search data

05 · Replication

Run it yourself in about ten minutes

SetupGoogle Trends · region United States · past 12 months · web search unless stated. Keep one constant anchor term in every comparison so separate charts stay comparable; note that each chart normalizes to its own peak, so levels are only meaningful within a chart.
Breakai consultant · ai integration · legal ai software · self-hosted ai · ai automation services — look at the week of 2026-07-12.
Test 1openai · google gemini · anthropic, run three times: web search, then YouTube search, then news search. Compare the same weeks across the three.
Test 2ai consultant · car insurance · divorce lawyer · plumber near me · quickbooks on web search. Watch whether the controls step down in the same week.
Test 3Switch any set to past 5 years and read June-to-September in each prior year before calling anything seasonal.
NoteTrends exports do not record the region, property or category you selected. Put them in the filename, or you will not be able to interpret your own file a week later.

06 · Limits

What this does not establish

Underlying exports, the full term list and the working are held by the author and available on request. Corrections are welcome; if you replicate this and get a different result, we would like to know.