Google's People also ask box is now written by AI. What changes for you?

The assumption in most content plans is that the question box is still a page contest: find the questions buyers ask, answer them plainly on your site, get quoted, get the click. That contest is over. It ended between May and August, and the businesses still paying for it will be the last to notice, because nothing in their reporting announces it.

What the two datasets say

Google started putting AI Overviews inside some People also ask answers in November 2024. In July 2025, Mark Williams-Cook at AlsoAsked measured the share at 12.6%, with the other 87.4% still "website-cited" — a snippet from one page, with a link (reported by Search Engine Roundtable, 3 July 2025). On 9 September 2026 the same tool, on a sample of 19.2 million English-language questions asked in 2026, put it at 97% for the first week of September, up from 86% in August. Saeed Khosravi at Allintitle, whose Also Ask Miner tracks the old page-quoting endpoint across a 6,092-question sample, reported its success rate falling from about 70% in November 2025 to zero from August 2026 — every answer now generated. Both are tool-vendor datasets, not a Google statement, and they agree with each other to within a few points. The month-by-month series from AlsoAsked's chart:

Source: AlsoAsked, 19.2 million English-language People also ask questions, 1 January to 7 September 2026, as published by Search Engine Roundtable on 9 September 2026. Vendor data; the trend is the point, not the decimal.

What it removes

The cheapest reliable route to a click that most mid-market companies ever had. For a decade the recipe was mechanical: list the questions your buyers type, write a page that answers each one in plain, literal sentences, and wait for Google to lift your sentence into the box with your link under it. Agencies sold it as a content strategy, and it was one — an honest, measurable one. I have no public figure for how much traffic that box sent any given company, and I won't invent one. What is measurable is the mechanism: the box no longer takes your sentence. It takes several sources, writes its own paragraph, and lists the names it drew on underneath.

That changes the economics of every FAQ page, every "questions we get asked" section and every question-headed blog post on your site. They are not worthless. They are now raw material for an answer somebody else writes, and the only thing your company gets out of it is whether it is named as a source — and whether the answer describes it accurately when it is.

What it rewards instead

An AI Overview is composed, not lifted. To write it, the model has to decide which companies it understands well enough to cite for this question, and it cites the ones it can describe in one confident sentence. That is a different test from "who answered the question best." I call the four conditions the Citable Entity Model: Identity — it is clear what the company is and does; Consistency — it is described the same way on its site, on LinkedIn, in directories and in the press; Corroboration — third parties say the same thing about it; Extractability — its pages state it in sentences a machine can quote. The old People also ask contest tested extractability alone. The new one tests all four, and extractability is the cheapest of them.

Which is why a company can answer every buyer question on its site and never appear in the box. The machine found the answer. It could not find the company — or found three contradictory versions of it and cited someone it could describe in one go. The mechanism is the same one behind why you rank on Google but AI never cites you; what has changed this summer is that it now governs the question box too, not only the AI answers at the top of the page.

Why the fix is upstream

The machine's questions are the buyer's questions from Reports, Not Revenue: why this company, for what exactly, compared to whom, with what proof. A company that cannot answer those in one sentence for a human cannot be summarised by a model either. The AI Overview is the screenshot wall run by software — thirty homepages offering "expertise and personal attention," and a system that cannot tell them apart, so it names none of them, or one of them at random. No amount of question-shaped content fixes that. It is a positioning decision first and a content task second, and the order matters, because the content task is cheap and the decision is the thing being avoided.

The practical sequence is the one I've set out before: make the pages legible, then make the company consistent, then earn corroboration. Structuring content so AI can surface and cite it covers the first; the signals that make a company citable covers the third. The one in the middle is the one nobody sells, because it is a decision, not a deliverable.

What to do this quarter

Do not commission "People also ask optimisation." The box that rewarded it has gone. Do this instead, and it costs an afternoon before it costs a euro: type the five questions your buyers actually ask into Google, expand the box, and read who it names and how it describes them. Then read how it describes you, if it does. That comparison — the machine's one-sentence version of your company against the one you'd write yourself — is the whole diagnosis. The method for doing it properly is in how to check whether you're discoverable in AI search.