Stop writing meta descriptions for every page

Meta descriptions have survived every prediction of their death, mostly because writing them is cheap and arguing about them is cheaper. The 2026 research, pulled together by Zoe Ashbridge at Search Engine Land on August 20, is the first body of evidence detailed enough to settle the practical question. The answer is not that meta descriptions are dead. It is that they matter enormously on a small number of pages and not at all on the rest, and almost every content team allocates its effort in exactly the wrong proportion.

Do meta descriptions matter to AI search engines

This is the newest question and it has the cleanest answer. Writesonic ran a test placing 62 hidden markers across pages and checking which ones surfaced in six AI assistants: ChatGPT, Claude, Gemini, DeepSeek, Grok and Copilot. Meta descriptions scored zero across every crawler tested. Not weak, not inconsistent. Zero. The full write-up sits in Search Engine Land's research round-up on meta descriptions, and it is worth reading in the original because the individual studies use different definitions of effect, which makes the consistency of the result more interesting rather than less.

Seer Interactive approached the same question from the other direction with a six-week experiment on 21 high-traffic pages, replacing the meta description with a period or a placeholder and watching what happened to AI crawl activity. The effect was a statistically flat 1.3% decline. Two studies, two methodologies, one conclusion: the field is not part of how generative engines decide what to read or what to cite.

That should not be surprising once you think about what a meta description is. It is a marketing summary written by the publisher, in the head of the document, aimed at a human deciding whether to click. An engine assembling an answer has the entire body of the page available and no reason to prefer the publisher's own promotional gloss on it. What gets cited is specific, quotable, verifiable claims in the body, which is the same pattern we found when documentation replaced community threads in ChatGPT citations and the same one behind the structure-level analysis of citation rates.

Do meta descriptions matter to Google's snippets

Here the answer is more uncomfortable, because it has been true for years and the industry has kept writing them anyway. Seer's analysis across thousands of keywords found Google displaying its own generated description rather than the supplied one for about 70% of queries. Individual pages showed between 2 and 11 different snippet variations in a single month. The snippet is not a field you set. It is a per-query decision Google makes, and your description is one candidate among several.

Search Engine Land ran the obvious test on itself, stripping meta descriptions from ten top articles for 15 days. Clicks went up by 18, or 1.8%. Impressions fell by around 6,000, or 5.88%. Click-through rate moved 0.1%, which is nothing. A small sample over a short window, and the authors are appropriately careful about it, but it is directionally consistent with the 70% rewrite rate: if Google is writing most of your snippets anyway, removing your version changes little.

The finding that complicates the easy answer

If the research stopped there, the advice would be simple and wrong: delete the field, move on. But the same body of work contains a result pointing hard in the other direction. Manually written meta descriptions delivered a 176% click-through lift over descriptions generated by ChatGPT-4, and beat Google's own rewrites by 54.1%. The AI-generated versions were not merely worse than human ones, they produced a 21.5% CTR drop in their own right.

Reconciling those numbers is the entire practical exercise, and it is not actually a contradiction. Google rewrites the description when the supplied one is generic, repetitive or a poor match for the query. A well-written, specific description that matches searcher intent is more likely to survive, and when it survives it converts better. The 70% rewrite rate is not evidence that descriptions do not work. It is partly a measure of how many of them are bad.

Which reframes the failure. The problem was never the field. The problem is that most organisations write meta descriptions the way they fill in alt text, as a compliance task assigned to whoever has capacity, judged by whether the box is populated rather than by whether the sentence would make anyone click. Fill 4,000 boxes with competent filler and you get the 70% rewrite rate, then conclude from it that the field does not matter, which produces more filler. That loop has been running for a decade. The way out of it is not a better template or a better generator. It is accepting that the field only pays when someone writes it who knows what the page is for and who it is competing against, which means it can only be done for a set of pages small enough that a person can hold them in their head.

“A field that is worthless when filled in mechanically and valuable when written deliberately is not a broken field. It is a field being used at the wrong scale.”

The economics of writing them by hand

The research puts a number on the cost side too, and it is bracingly high: 19.1 minutes per page for a human-written description versus 2.5 minutes with AI assistance. Round it to twenty minutes. For a 5,000-page site, writing every description by hand is roughly 1,667 hours, which no one is going to authorise and no one should.

But run the same arithmetic on a realistic target set. Most sites have somewhere between 50 and 200 pages that carry the overwhelming majority of commercially valuable impressions. At twenty minutes each, 150 pages is 50 hours. One person, a bit over a week, and the 176% differential applies to precisely the pages where click-through translates into revenue. That is one of the better returns available in on-page work, and it is available right now without a tooling decision or a migration. It also has an unusually short feedback loop for search work. Snippet changes surface within days rather than the months a content or link investment takes to resolve, so a team can test the approach on twenty pages, look at the click-through data three weeks later, and decide whether to fund the remaining hundred and thirty on evidence rather than on faith.

Leaving the tail empty is a real option and it deserves defending, because it makes people nervous. If Google writes its own description for 70% of queries, and the ones it writes for your low-value pages are drawn from body copy that is probably more query-relevant than a template would be, then an empty field on page 4,000 costs you close to nothing and saves the effort. The audit tools will flag it. The audit tools are counting completeness, not value. This is the same pattern as treating AI visibility gains as technical debt repayment: the tool reports what it can count, and what it can count is not what matters.

A triage rule for the next hundred pages

Sort every indexed page by impressions over the last 90 days, then apply three tiers. It takes an afternoon to build the list and it replaces a policy most teams have never actually written down.

That fourth card is the one people skip and it is the one that compounds. Checking whether your description survived, on the twenty pages you care most about, tells you something no tool reports: which of your descriptions Google judged worse than what it could write itself. That is direct feedback on your writing, from the only evaluator whose opinion decides the outcome, and it costs about ten minutes a month to collect.

So the plan for this week. Pull the impressions report, cut it at the point where the numbers stop being commercially interesting, and take that list to whoever owns on-page work. Everything above the line gets written by a person. Everything below it gets automated or ignored, and the audit warning gets muted rather than resolved. If you want the fuller argument for optimizing the small set of pages that carry the weight, that is what our on-page optimization engagements are built around, and it pairs with the citation-side work we described in treating AI answers as a bibliography.

See where you are cited today

A free snapshot audit of your rankings and AI citations before we ever talk.