Meta Descriptions in AI Search: Does Anything Still Read Them?

By Muntazir · Published 20 October 2026

In 2020, Ahrefs ran 20,000 keywords through Google and compared the meta description each page’s owner had written against the snippet Google actually showed. The two matched 37% of the time. For the other 62.78%, Google had quietly substituted its own excerpt from the page — more often for long-tail queries (65.62%) than for head terms (59.65%). A separate study by Portent, looking only at first-page rankings, put the rewrite rate even higher: 71% on mobile, 68% on desktop.

Most of what gets written into a <meta name="description"> tag, in other words, is a draft Google is free to ignore.

Why would Google ignore something a site owner wrote on purpose?

Google’s own documentation on search snippets states the logic plainly: snippets are “primarily created from the page content itself,” and the meta description comes into play only when Google decides it “might give users a more accurate description of the page” than anything else on it.

That’s not a dismissal of the tag — it’s a design choice with a real constraint behind it. A meta description is written once and attached to a URL. A search query is different every time. Someone looking for “return policy” and someone looking for “shipping cost” might land on the same page, and a single sentence written in advance can’t be the best possible answer to both. Google’s snippet system looks at the actual query and pulls whichever passage of the page answers it, which explains why the rewrite rate climbs for long-tail searches: the more specific the question, the less likely one fixed description happens to cover it.

So what does a meta description actually control?

The fallback case — the roughly 37% Ahrefs found unchanged, and whatever the true figure is today. A page’s owner doesn’t get to pick when that fallback kicks in; only the content and a well-written description determine whether it’s good enough to count.

A technical audit that reads a page’s markup — the same category of check behind a free on-page SEO review — looks for two specific failures here. The first is a missing description altogether — not a bad one, just none — which leaves a search engine or an AI summary with nothing written by the page owner to start from. The second is length: a homepage description long enough that it gets cut off mid-sentence in the result, typically past somewhere around 155 to 165 characters depending on the surface showing it. Neither finding says the description will be used. Both say what happens if it has to be.

Does any of this reach an AI-generated answer?

Here the evidence runs out earlier than the question does. Google has stated clearly that the nosnippet and max-snippet robots directives — which govern ordinary snippet text — also apply to AI Overviews and AI Mode: a page marked nosnippet doesn’t just lose its blue-link snippet, it’s excluded as a direct input to those two generated-answer features as well. That’s a real, documented link between snippet-level controls and AI answers.

What Google hasn’t stated is narrower and more specific: whether the <meta name="description"> tag itself, as opposed to the snippet system built around it, is something an AI Overview or AI Mode reads as an input. The directive that controls exposure and the tag that sometimes supplies the text are related but not identical, and the documentation that’s explicit about the first says nothing about the second.

OpenAI’s side is quieter still. Its crawler documentation is detailed about access control — which bot fetches what, how fast a robots.txt change takes effect — and says nothing about how a fetched page’s meta tags factor into the text ChatGPT eventually produces. That’s consistent with how a language model actually works: it isn’t picking a pre-written sentence to display the way a search results page does. It reads rendered content and generates new text to answer the question asked. A sentence written in advance for a different purpose, aimed at a different audience, written to be clicked rather than read for understanding, has no obvious place in that pipeline — which is a reasonable guess, not something either company has confirmed or ruled out in writing.

What’s actually worth doing about it

Writing one good description per important page remains the only thing a site owner controls directly, even with the Google figures as they are. It doesn’t guarantee the sentence appears anywhere, in a search result or an AI answer. It guarantees that when something does fall back to it — and roughly a third of the time, at least on Google’s own 2020 numbers, something does — there’s a real sentence there instead of a blank field.

A free SERP preview shows what a page’s current description looks like as a search snippet today. It won’t show what an AI Overview or a chatbot does with the same page, because that pipeline doesn’t expose a preview the way a search results page does — which is the honest shape of the gap this whole question runs into, and the same gap that runs through most of technical AI search work: confirming a page can be read is one project, confirming what a model does with it afterward is a different one, and the second is still mostly unobservable from outside.