Blog · Foundations

Does SEO still matter for AI search

Published 6 October 2026 · Updated 12 October 2026 · 9 min read

Does SEO still matter for AI search
Photograph: Alexander Andrews / unsplash

The short answer: Yes, but only the half of it that gets you into the candidate pool. Crawl access, server-rendered HTML and a live index entry are what put a page in front of a retrieval system at all. Rank position, on the other hand, has stopped being a reliable predictor of whether you get cited. Those two facts pull in opposite directions, and most teams are still spending as though only the first one exists.

The useful version of this question is not "is SEO dead". It is narrower and more answerable: which specific SEO inputs still move an AI answer, and which ones now move nothing at all.

The index is still the gate, and that part has not moved

Google's own optimisation guidance for its AI features is blunt about this. Its guide, last updated on 10 July 2026, says there is no special markup to add, and that you do not need to create new machine readable files, AI text files or Markdown to appear in Search. For Google, the eligibility rules are the ordinary Search ones, because that guide says its generative features draw on the Search index. So if robots.txt blocks Googlebot, if the content only exists after JavaScript runs, or if the page is orphaned with no internal link pointing at it, you are not in the pool Google draws from. Other engines publish much less about how they retrieve, and we have not found equivalent documentation for all of them, so treat the same reasoning elsewhere as a sensible default rather than a documented rule.

So the boring technical layer did not become optional. It became load bearing in a way it never was before, because there is no second chance further down the page. A link at position fourteen used to collect a trickle of clicks. A document that was never retrieved collects nothing, and leaves no trace you can diagnose from a rank tracker.

The same guide describes query fan-out, which is how one typed question becomes several concurrent machine-generated queries. Your page is not competing for the question the user asked. It is competing for four or five reformulations of it that you never see. That is also the mechanism behind why ranking is the wrong frame for AI search.

Where rank stops predicting citation

Here is the number that should change how you budget. A measurement study of 55,393 trending queries, issued between 13 March and 21 April 2026 by researchers at Washington University in St. Louis, found that across 7,562 AI Overviews only 48.0 per cent of cited domains also appeared in the top ten results on the same page. Nearly three in ten cited domains, 29.8 per cent, appeared nowhere on the first page at all. Both figures describe Google AI Overviews on trending queries over a 40-day window, not AI search in general.

Read that twice. On that sample, slightly more than half of the cited domains were ones Google's own ranking, on the very same query, had not placed in its top ten. The citation set was also less concentrated: the five most cited hostnames accounted for 20.0 per cent of AI Overview citations, against 47.0 per cent of the first page links in the same comparison.

What does that imply for a mid-sized brand? Two things. Holding a top three position is no longer a moat, because the engine is drawing from a wider and flatter pool than the one you fought for. And sitting at position twenty-two is no longer a sentence, because retrieval can reach you there if your page carries the specific passage the fan-out query needed. On the evidence of that one dataset, topical specificity looks more valuable relative to domain strength than a rank-first model would predict. We have not seen a comparable public dataset testing the same thing on another engine.

The failure mode SEO never had to think about

Ranking has one output: a position. A citation has two. You are named, and a claim is attached to your name. The second one can be wrong while the first one looks like a win.

The same 2026 study scored 98,020 atomic claims against the pages they cited. 11.0 per cent were unsupported by the source: 6.98 per cent were simply not addressed anywhere in the cited page, 2.66 per cent were contradicted by it, and 1.39 per cent were ambiguous.

A contradicted claim is the one to care about. It means an engine cited your page and then said something your page disputes. A rank tracker will not surface that, because it records a position and not the sentence attached to your name. A traffic report usually will not either, because the user who read the wrong sentence often does not click. That is why we think the honest unit of measurement is a sampled, repeated read of the actual answer text, which is what an audit produces and what a single dashboard score cannot.

Ambiguity in your own copy is the cheapest contributor to fix. If one page says a week, another says seven working days and a third says five, a summariser picks one and states it flatly. Reconciling your own numbers across your own site is cheap, and in that study omission and contradiction, rather than authority, were the dominant failure modes. We cannot show you a controlled comparison against link building. We can only say which failure the published data actually found.

The click arithmetic underneath all of this

Pew Research Center tracked the browsing of 900 US adults across 68,879 Google searches in March 2025. When an AI summary was present, users clicked a traditional result in 8 per cent of visits. Without a summary, 15 per cent. Links inside the AI summary itself were clicked in about 1 per cent of visits. Pew describes this as observational panel data rather than an experiment, so on its own it shows association and not cause.

A preregistered randomised field experiment published in August 2026, with 1,100 participants whose Google interface was manipulated by a browser extension for seven days, measured a drop of 18.8 percentage points in click-through when participants were pushed into AI Mode, and a rise of 8.8 percentage points when AI Overviews were removed. That is a causal estimate rather than a correlation drawn from a traffic chart. It also covers one engine over seven days with 1,100 participants, so read it as short-run evidence about Google Search.

The practical reading is not that traffic has vanished. Two sites with identical rankings can now have very different outcomes, and the difference is decided after the ranking, inside the answer.

A check you can run this week

Pick fifteen questions your buyers genuinely ask. Not head terms. Real sentences, the kind with a verb in them, since Pew's panel analysis found that query length, a leading question word, and the presence of both a noun and a verb each raise the chance that a Google AI Overview is generated at all.

For each one, record three columns. Where do you rank organically. Were you cited in the AI answer. Was the claim attached to your name correct. Run each question at least three times across two engines, on different days, because a single run tells you almost nothing.

Then sort the table. The rows where you rank well and are not cited are retrieval or passage problems, and they usually trace back to page structure or to a missing explicit statement of the fact. Rows where you are cited but ranked poorly tell you what to write more of. Rows where the claim was wrong are the urgent ones. The self-check walks through ten of these tests in more detail if you want a starting script.

Two inputs make the retrieval rows easier to fix. Sitemaps, feeds and freshness signals that matter to AI govern how fast a corrected page is seen, and structured data for AI search governs whether your identity resolves cleanly even though the markup is not itself a ranking input. None of this is new work invented for a new acronym. It is the same discipline, pointed at a different output, which is the whole argument in what answer engine optimisation is.

Honest limits

This post does not establish that any of it holds outside Google. Every citation, overlap and concentration figure above comes from Google AI Overviews on trending queries in a single 40-day window, and the click figures come from Google Search panels and experiments, so commercial queries on ChatGPT, Perplexity, Copilot or Gemini may behave differently and we have not found a comparable public dataset for them. It also does not establish that topical specificity causes citation; the study measures overlap, not mechanism. Attribution stays unsolved, because most of these surfaces send no usable referrer we can read, so we cannot connect a mention to a sale. Durability is unknown too: a fix that restored citations in March may not hold in September, since the retrieval stack changes without announcement. We re-run our own prompt sets monthly for that reason, and those runs are a sample of a handful of engines on a scoped prompt set, never a census of what an engine says.

FAQ

Should I stop doing technical SEO? No. It became the entry condition rather than the competitive edge. Crawl access, server rendering, internal linking and index coverage decide whether you are in the candidate set, and nothing downstream can rescue a document that was never retrieved.

If rank does not predict citation, why rank at all? Because the correlation is weakened, not absent, and ranking still brings the clicks that do happen. A 2026 measurement study of trending queries put the share of AI Overview citations that also sit in Google's own top ten at just under half, 48.0 per cent. That is a meaningful head start, just not a lock.

Does adding more schema get me cited? Google states directly that structured data is not required for its generative features and that no special markup exists for them. Markup still earns rich results and still helps an engine resolve who you are, so keep it. Treat it as identity plumbing, not as a lever.

How do I find out that an engine is saying something false about us? Only by reading the answers. Sample the questions that matter, log the sentence next to your brand, and re-run on a schedule. The sample audit report shows how we lay that evidence out so a finding can be argued with.

Is this different for a small site with no authority? On the Google AI Overview data above it tends to help you, since the citation pool measured there was flatter than the organic one, and a page that states one specific fact well can be retrieved for a fan-out query that a larger, vaguer site never addressed. We have not seen the same measured on other engines.

Sources

Related reading

What is answer engine optimisationWhy ranking is the wrong frame for AI searchSitemaps, feeds and freshness signals that matter to AIStructured data for AI search: what to mark up firstHow an AI answer is assembled, step by stepHow long AEO takes to show up in answers