People Also Ask Is Now 97% AI. The Old Answer Endpoint Is Gone.

Google's "People Also Ask" box is no longer a snippet box. Two independent datasets put the share of PAA answers that are AI-generated at 97% in the first week of September 2026 — and one of them argues the real figure is 100%, because the server endpoint that returned the old single-source answers stopped responding in August. Fourteen months earlier, that share was 12.6%.

That is the whole story in three numbers. The rest of this piece is about the part that actually costs you something: the old PAA answer was one link to one page, and the new one is a generated paragraph with several citations underneath it.

Key takeaways

• 12.6% → 97% in fourteen months. A July 2025 analysis of 8.4 million PAA results found 12.6% were AI-generated. A September 2026 sample of 19.2 million English-language answers found 97%.
• 86% to 97% happened in a single month — August to the first week of September 2026.
• A second dataset says 100%. Allintitle's sample of roughly 6,000 PAA questions, drawn from saved raw server responses between November 2025 and September 2026, shows the legacy answer endpoint returning nothing at all from August 2026 onward.
• The endpoint died gradually, then suddenly: roughly a 70% success rate in May 2026, 25.8% by the end of July, zero in August.
• The format changed, not just the generator. PAA answers moved from a single-source extracted snippet to a generated answer carrying multiple source citations.

How much of People Also Ask is AI-generated now?

Here is the trajectory, with each figure tied to the sample it came from. Sample sizes matter here, and most write-ups omit them.

Two things are worth noticing. First, the curve is not linear — it moved slowly for a year and then jumped eleven points in roughly four weeks. Second, the two organisations measuring it are measuring different things, which is why they disagree.

Why do the two datasets disagree — 97% or 100%?

Because one counts answers and the other counts responses.

AlsoAsked samples rendered PAA answers at enormous scale — 19.2 million of them — and classifies each as AI-generated or not. At that volume, a 3% residue can be almost anything: cached results, regional variation, classification error, or a genuine tail of old-style answers still being served.

Allintitle went the other way: a much smaller sample of about 6,000 questions, but inspecting the saved raw server responses behind the calls rather than the rendered box. From that angle there is no 3% residue to argue about, because the legacy endpoint simply stops returning data in August 2026.

Neither number is wrong. The honest reading is that Google retired the old PAA answer mechanism during August 2026, and the 3% gap is measurement lag rather than a surviving population of classic snippets. If you need one figure for a slide, use 97% and note that the underlying mechanism appears to be fully retired.

What actually changed inside People Also Ask?

The share statistic gets quoted; the format change is the one that changes your work.

Before: a PAA answer was an extracted snippet. One passage, from one page, with one link. Winning that slot was a binary outcome — your page was the answer, or it wasn't. It behaved like a featured snippet, and you could optimise for it the same way.

After: a PAA answer is generated text with multiple source citations attached. There is no single winner. Several pages contribute, each gets a citation, and no single citation carries the weight the old link did.

That is a structural change in how attribution is distributed, and it points the same direction as everything else happening on the results page. We covered the parallel case in Google Is Shrinking AI Overview Citations — more citation slots, each less visible, with click-through on AI Overview citations already sitting around 1%.

Google's own documentation has not caught up with any of this. The visual elements gallery still describes PAA as a "related questions group" — "a cluster of questions that are related to what the user initially searched for" — and makes no mention of AI Overviews anywhere on the page. It does, however, state the position plainly: "While you can't control what shows up here, it can be helpful to pay attention to the related search queries when you're thinking about topics you could write about."

That sentence was written about the old PAA. It is more true of the new one.

What breaks in your reporting?

This is the part nobody is talking about, and it is the part that will make a report wrong.

If you track PAA presence — in-house scripts, a rank tracker, a content-gap tool, a spreadsheet somebody maintains — that data has a discontinuity in August 2026. Whatever was reading the legacy endpoint stopped getting answers. Depending on how the tool was written, you are now looking at one of three failure modes:

Before/after comparisons that straddle August 2026 are not comparable. If you are building a quarterly report right now, annotate the break rather than explaining the shape of the curve as a performance change. And if a client asks why their PAA visibility "fell off a cliff," the first thing to check is whether the feature they were being measured against still exists in the form the tool assumed.

Deciding which of your AI-search metrics still mean what they used to is its own exercise — our AI search KPI guide walks through which numbers survive this kind of change and which quietly stop being comparable.

Can you opt out of AI-generated PAA answers?

Not selectively. Google's AI features documentation is explicit on both halves of this.

On getting in: "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary." Standard eligibility applies — indexed, snippet-eligible, meeting technical requirements and Search policies. And a caveat worth reading twice: meeting every requirement "doesn't mean that Google will crawl, index, or serve its content."

On staying out, the available controls are nosnippet, data-nosnippet, max-snippet and noindex. Google calls these "the control for site owners to manage access."

Note what that list does not contain: anything scoped to AI features. These are the same preview controls that govern ordinary snippets. Suppressing your content in generated PAA answers means suppressing your snippets everywhere — in classic results, in featured snippets, across the whole preview surface. There is no dial marked "AI only."

data-nosnippet is the one exception worth knowing about, because it is per-element rather than per-page. You can wrap a specific block — a pricing figure, a proprietary definition, a paywalled excerpt — and leave the rest of the page fully eligible. That is a scalpel where the others are a hammer, and it is the only surgical option Google actually gives you.

If you are unsure what your pages currently declare, our meta tags checker will show you the robots and preview directives you are actually emitting, which is regularly not what the CMS settings screen claims.

What should you actually do about it?

The strategic shift is small but real: PAA used to reward being the answer, and now rewards being citable. Those are not the same target. The first favours a page that owns a question outright. The second favours a page that offers several clean, quotable, factually checkable passages a generated answer can lift from — which is much closer to how you would already write for AI Overviews and AI Mode.

If you want a fuller treatment of that second target, how to get cited by AI covers the structural patterns that make a passage liftable.

Frequently asked questions

Are all Google People Also Ask answers AI-generated now?

Effectively, yes. AlsoAsked measured 97% across 19.2 million English-language answers in the first week of September 2026. Allintitle's inspection of raw server responses puts it at 100% from August 2026, on the grounds that Google's legacy PAA answer endpoint stopped returning data entirely that month.

When did Google switch People Also Ask to AI answers?

The switch completed in August 2026, but it was gradual. The legacy answer endpoint had roughly a 70% success rate in May 2026, fell to 25.8% by the end of July, and returned nothing from August onward. The AI-generated share itself moved from 86% in August to 97% in the first week of September 2026.

Did Google announce the People Also Ask change?

No. Google's search documentation still describes People Also Ask as a "related questions group" and does not mention AI Overviews in that context at all. The change was identified by third-party datasets, not by an announcement.

Can I stop my content appearing in AI-generated PAA answers?

Only with the standard preview controls — nosnippet, data-nosnippet, max-snippet and noindex. None of them are scoped to AI features, so page-level use removes you from ordinary snippets too. data-nosnippet is the only per-element option and the only way to protect a specific block without sacrificing the whole page.

Does winning a People Also Ask slot still send traffic?

Less, and differently. The old PAA answer linked to a single source. The new one attaches multiple citations to generated text, so attribution is split across several pages and no single link carries the weight it used to.

Should I still write content targeting PAA questions?

Yes, but for a different reason. The questions themselves remain a genuine signal of what users ask. What has changed is the payoff: you are no longer competing to own one slot, you are competing to be one of the sources a generated answer cites.

Where this leaves you

The 97% figure is the headline, but the useful part is smaller and more practical: an entire measurement surface changed shape in August 2026, and most reporting has not accounted for it. Check your data for the break before you try to explain the curve.

Then check the part that is still under your control. Start with whether AI crawlers can reach your pages at all — our free crawler checker tests that in a few seconds — and confirm your preview directives with the meta tags checker. Neither will get you into a generated answer on its own, but both will tell you whether you have accidentally ruled yourself out.