Blog · Foundations

Why AI search is not a ranking game

Published 20 February 2026 · Updated 12 October 2026 · 9 min read

Why AI search is not a ranking game
Photograph: Zoshua Colah / unsplash

The short answer: There is no list, so there is no position to hold. Google documents its own AI features turning one question into several related queries, retrieving results for each, and writing a single answer from what comes back. You are not competing for slot three. You are competing to be the passage that survives selection, often for a query the buyer never typed. How far that holds on the engines that publish no such description is inference, not observation.

Rank is a scalar. One query, one list, one number you can watch move. Everything about the old toolchain, the weekly report, the position graph, the keyword cell in a spreadsheet, depends on that list existing. In an answer surface it does not.

One question becomes several, and you never see them

Google documents the mechanism by name, for its own features. Query fan-out, in its guide to optimising for generative AI features, is a set of concurrent related queries the model generates to request more information and fetch extra results beyond the literal question. Its Search Console documentation describes AI Mode the same way, grouping the question into subtopics and searching each one at the same time. Both pages were open and said this on 12 October 2026.

Only Google describes this in public. Of the five engines we audit, it is the one whose documentation we can quote on the point, so treat fan-out elsewhere as a reasonable inference from the sources answers cite, not as documented behaviour.

Sit with what that does to a ranking report. The buyer typed one thing. The retrieval layer ran five or ten things, each with its own candidate set, each pulling pages that would never have surfaced for the original phrasing. Your competitor can lose the typed query outright and still own three of the sub-queries that actually fed the answer.

Which is why keyword-level position tells you so little here. You can hold the top slot for the phrase and be absent from the paragraph, because the paragraph was assembled from questions nobody showed you. Working out what those hidden questions are is most of the job that What is answer engine optimisation describes.

Selection happens at the passage, not the page

A ranked list orders documents. A generated answer stitches fragments.

Retrieval-augmented systems, as the published research on them describes, chunk what they fetched, score the chunks, and discard most of them. What reaches the model is a short set of passages, stripped of your navigation, your preamble and the paragraph three screens down where you finally answered the question. The model then writes from that residue. Not one of the five engines we audit publishes its chunking or scoring, so the shape is taken from the research literature and from which of our pages get quoted, not from a vendor specification.

So a page can be excellent and still fail, because its best sentence never travelled alone. A heading that promises and a body that meanders produces chunks that contain a promise and no answer. Put the claim, the number and the qualifier inside one self-contained passage and you change what is available to be lifted. That is a formatting decision with a retrieval consequence, not a style preference.

The peer-reviewed work here is modest and worth reading honestly. The paper GEO: Generative Engine Optimization, accepted to KDD 2024, reports that content-level changes such as adding citations, quotations and statistics boosted visibility by up to 40 percent on GEO-bench, the benchmark the authors built, and says plainly that the effect varies across domains. That figure is from the abstract, read on 12 October 2026. It is a 2024 benchmark result on one set of queries, not a promise about your category or about what the engines do today.

Google says there is no separate AI ranking factor

This part surprises people who have been sold a factor list.

Google's own guidance on optimising for generative AI features says its AI features are rooted in its core Search ranking systems, and that standard practice continues to apply. It also states plainly that structured data is not required for generative AI search and that there is no special schema.org markup you need to add. Both sentences were on the page on 12 October 2026, and both are inconvenient for anyone selling a separate AI checklist.

That is Google speaking about Google. The other engines publish nothing equivalent, so the honest reading is that one engine has denied a separate factor list and the rest have not been asked in public.

Read carefully, that is not an argument for doing nothing. It means the levers are eligibility levers rather than scoring levers. Can the fetcher reach the page, parse it, and resolve who published it. Markup still earns its place for that last part, which is the honest reason to care about structured data for AI search: disambiguation and rich-result eligibility, not a hidden bonus in an answer-ranking formula that nobody has published.

The failure mode: reading a rank report for a surface with no ranks

Here is the specific way teams get this wrong in 2026.

Google mixes AI Mode and AI Overview data into Search Console's Performance report, inside the Web search type. The help page on impressions, position and clicks sets out the counting rules, checked 12 October 2026: a click to an external page in AI Mode counts as a click, standard impression rules apply, an AI Overview occupies a single position with every link in it assigned that position, and a follow-up question inside AI Mode is treated as a new query whose impression, position and click data belong to that new query. A separate Generative AI performance report, which Google says reached all sites on 31 August 2026, reports impressions for AI Overviews and AI Mode on their own.

Each of those rules is reasonable. Together they produce a number that cannot be read the old way. Impressions arrive from surfaces where no list existed. Average position blends a list coordinate with a non-list coordinate. Click-through rate falls because answers resolve without a visit, which looks exactly like a ranking decline on a chart and is not one.

We have watched teams respond to that chart by rewriting pages that were performing fine. That is an observation from the client work we have done this year, not a survey, and we have not counted how often it happens. If your average position drifted and your impressions rose through 2026, check the date against the reporting change before you blame the content.

A related trap is treating a single run as evidence. In our own audit re-runs, the method being a set of roughly 45 prompts put to five engines with retrieval on and retrieval off, the same prompt minutes apart has come back citing different pages. We have not published a run count or a stability rate for that, so read it as a sample of our own work rather than a measured property of the engines themselves. Either way, a claim about your visibility built on one screenshot is an anecdote with good lighting.

What to measure instead

Replace the position number with a frequency. It is less tidy and far more honest.

  1. Write twenty to forty questions a buyer would actually ask, in their words, including

    2. the comparison and objection ones you would rather they did not ask.

  2. Run each question on more than one engine, at least five times, on different days.
  3. Record three fields per run: were you named, was the claim about you correct, which

    4. sources did the answer lean on.

  4. Compute a mention rate per question, not an average across the set. Averages hide the

    5. two questions that matter.

  5. Keep the cited URLs. The pattern in other people's cited pages tells you what the

    6. retrieval layer currently rewards in your category.

Then change one thing and re-run. Unblock the fetcher, wait a crawl cycle, re-run. Rewrite the buried answer into a self-contained passage, wait, re-run. Slow and checkable beats a score that moves for reasons nobody can name.

Two more things fall out of this framing. Mention rate decays, so refresh is a maintenance schedule rather than a project, which is the subject of how often to refresh content for AI search. And a mention inside someone else's answer is not an asset you hold, which is the whole point of the rented versus owned split. An audit is the compressed version of this measurement: roughly 45 prompts across five engines, run with retrieval on and retrieval off, with the evidence attached to each finding so it can be argued with. The sample audit report shows the shape the report takes.

Honest limits

Three gaps, stated plainly, and one more besides.

We cannot see the fan-out queries directly. We infer them from cited sources and from what the answer covers, and inference is not observation. Second, attribution from a mention to revenue stays weak: in the client logs we have looked at, referrer strings from these surfaces are mostly missing or stripped, which tells us what we cannot see rather than what did not happen. Third, we do not know the half-life of a fix. Our own re-runs move month to month, we have not measured that drift to a number, and anyone quoting you a durability figure is quoting a guess.

We have also not tested the claims above on engines that publish no documentation, our passage-level account of selection comes from research and from which pages get quoted rather than from any vendor specification, and the 40 percent figure belongs to a 2024 benchmark and to nothing we ran.

FAQ

Do rankings still matter at all for AI search? They matter as a proxy for eligibility. Google states that a page must be indexed and eligible to be shown in Search before it can appear in its generative AI features, and that those features are rooted in its core ranking systems. We have no openable source for how closely AI citations track the top ten, so we are not going to put a number on it. Treat ranking as necessary groundwork rather than as the scoreboard.

If there is no position, what do I report to my board? Mention rate on a fixed question set, correctness of the claim attached to the mention, and the share of answers citing a source you control. Report the question list alongside the numbers so nobody can quietly change it later.

Why does the same prompt give different answers? Retrieval is re-run per request and the candidate set shifts, and sampling in the model adds more variation. We see it in our own audit re-runs rather than in anyone's documentation. Repeat runs on different days are the defence.

Is this just SEO with new words? The plumbing overlaps and the goal does not. Crawlability, clean markup and real content serve both. The target changes from a coordinate in a list to a sentence that names you correctly.

Sources

All four were opened and checked on 12 October 2026.

Related reading

What is answer engine optimisationRented visibility vs owned visibility in AI searchStructured data for AI search: what to mark up firstHow often to refresh content for AI searchAEO vs GEO vs SEO: what the three terms actually meanAn AI search glossary that defines the terms properly