AI Overviews: what changed, and what to do about the clicks

Why this one is different from the assistants

The other pieces in this cluster deal with systems that retrieve independently and cite a handful of sources. AI Overviews is not that. It sits inside Google’s results page, draws on Google’s own index and ranking, and appears above the links you were already competing for.

Two consequences follow, and both make it a sharper problem than chat assistants.

• It affects traffic you already have. A chat assistant not citing you costs a click you never had. An Overview appearing above your result takes clicks you were receiving last quarter.
• The path to appearing in one is conventional. Sources are drawn largely from pages already ranking well for the query, so there is no separate optimisation route. You rank, or you are not considered.

That second point is genuinely good news, in that it means most of the work is work you were already doing. The broader picture of how these systems differ is in the LLM SEO overview.

Where the clicks actually go

The damage is not evenly spread, and the shape of it is more useful than any headline percentage.

Queries that can be answered completely in a paragraph lose the most, because the Overview satisfies the need entirely. Queries where the answer depends on the person, or where the next step is a purchase, lose far less.

Query type	Click impact	Why
Definitional, "what is X"	Severe	Answered completely above the fold. There is no reason to click
Simple how-to, single step	Severe	Same reason
Comparison, "X vs Y"	Moderate	Summarised, but buyers usually want detail and sourcing
Commercial, "best X for Y"	Light	People want to see options and pricing themselves
Navigational or branded	Negligible	The person already knows where they are going

The pattern holds across most reports, though the magnitudes vary widely

The strategic reading of that table: informational top-of-funnel content has become materially less valuable, and commercial-intent content has become relatively more so. That is a shift in where to invest rather than an argument for panic.

It also sharpens an argument the site makes elsewhere. A programme built on high-volume informational guides was always weaker than it looked in a traffic report, and this makes the weakness visible.

Being cited in an Overview

Two conditions, and the first is non-negotiable.

You have to rank. Sources are drawn overwhelmingly from pages already performing well for the query, so everything conventional still applies and there is no shortcut around it.

Then the passage has to be extractable. This is the same requirement as any retrieval system: a section that answers one question completely, states its claim before its qualifications, and names its subject rather than referring back to it.

01

Rank in the top results

The prerequisite. Nothing else matters if you are not in the consideration set, and the consideration set is conventional.

02

Answer the question in the opening lines

The extractable unit is usually a short passage near a heading that matches the query. Burying the answer in paragraph four removes you.

03

Be specific enough to be worth quoting

A number, a named method, a concrete constraint. Generic phrasing is interchangeable, and interchangeable sources get swapped out.

The practical detail worth acting on: match a heading to the question as a searcher phrases it, then answer it immediately beneath. That single structural habit does more here than any amount of markup.

The underlying reason is the same one that governs every retrieval system: the extractable unit is a passage rather than a page, which what AI search actually retrieves from a page sets out in detail.

Worth ruling out one thing before concluding your content is the problem. Contradictory structured data, especially two blocks claiming the same identifier, produces a page whose own description of itself disagrees, and the Schema Validator checks a whole page at once rather than a block at a time, which is the only way to see it.

What being cited is worth

Less than people assume, and worth being clear-eyed about before optimising hard for it.

A citation in an Overview produces a link, and links inside Overviews are clicked at a low rate because the answer is already there. So the realistic value is visibility and credibility rather than traffic: your name appears at the moment of the question, and some proportion of people will remember it or search you later.

Low

click-through, because the answer is above it

Real

brand exposure at the moment of the question

Zero

guarantee of persistence, they change constantly)

Which is why chasing citations on informational queries has a poor return

This is the awkward conclusion of the piece. Optimising heavily to be cited on a query that used to send you clicks will not restore those clicks, because the mechanism that took them is the same mechanism that would cite you.

So the honest response on badly affected informational queries is often not to fight harder. It is to accept the query is answered, stop investing in it, and move the effort to queries where a person still needs to see options, pricing or detail.

The queries worth deliberately conceding

This is the part most writing on the subject avoids, because conceding sounds like defeat. It is usually the correct call and it frees real budget.

Some queries are now answered completely on the results page and will stay that way. A definitional question with a stable, uncontested answer has no remaining reason to send a click, and no amount of rewriting changes that. Continuing to invest there is spending against a mechanism rather than a competitor.

Signal	Reading	Action
Answer is one uncontested paragraph	Permanently absorbed	Concede. Keep the page if it supports a cluster, stop investing
Answer depends on the reader’s situation	Not fully answerable	Keep. The Overview will summarise and people will still click
Answer changes over time	Recency favours you	Keep, and update it visibly
Query has purchase intent	Largely unaffected	Invest here instead. This is where the freed budget goes

The first row is the one teams resist and the one that saves the most

Conceding does not mean deleting. A definitional page that no longer earns clicks may still be the piece that establishes a term the rest of your cluster depends on, and removing it takes its internal links with it. Stop investing, keep the asset.

The freed capacity is the point. A quarter previously spent producing informational guides at the top of the funnel buys a good deal of comparison and pricing content, which is where the clicks now are and where they convert better anyway.

The response that actually works

Given all of the above, the useful reaction has three parts and only one of them is about Overviews.

Find out what you actually lost

Split Search Console by query type before concluding anything. Most sites find the damage concentrated in a small set of definitional queries and almost absent elsewhere, which is a much smaller problem than the headlines suggest.

Move investment toward commercial intent

Comparison pages, pricing explanations, alternatives pages. These lose the least, convert the most, and were usually under-invested relative to guides in the first place.

Make the surviving informational pages extractable

For queries you still want, heading matched to the question and the answer immediately beneath. Cheap, and it improves the page for readers regardless.

What not to do: rewrite everything, remove content because it lost clicks, or start a project. A page that lost half its traffic and still converts is not a failed page, and deleting informational content that supports a cluster damages the pages it was linking to.

The audit method for deciding what to keep, update or remove is set out in the content audit piece, and it applies unchanged here: judge a page on what it earns, not on what it used to earn.

What this does to the case for a content programme

The uncomfortable question underneath all of this: if informational queries lose their clicks, does the standard content playbook still work?

Partly, and with a changed shape. The playbook that no longer works is the volume one, where a large library of guides on high-volume questions produces traffic that is assumed to convert eventually. That model was always weaker than its traffic reports suggested, and Overviews have removed the traffic that disguised it.

What still works is narrower and was always the better version. Content that answers questions a general summary cannot: things specific to a situation, backed by data nobody else has, or making a judgement rather than reciting a definition.

Original data or research very defensible

Judgement and opinion with reasoning defensible

Situation-dependent guidance moderately

Synthesis of existing sources poorly

Definitional explainers barely

The bottom two rows are most of what most content programmes produce

The bottom two rows describe the majority of B2B content published, which is why this feels like an existential change to some teams and barely registers with others. If your library is mostly synthesis, a large fraction of it has just lost its rationale.

The constructive reading: this pushes content toward being genuinely worth reading, because the alternative no longer produces traffic. That is a harder standard and a better one, and it is the same conclusion the AI-search work arrives at from a different direction.

What to watch, and what to ignore

Overviews appear inconsistently. The same query can show one today and not tomorrow, and coverage varies by country, device and how commercial the query looks.

That instability means the useful monitoring is coarse rather than precise. Track whether your important queries tend to trigger one, and track the click-through rate on the queries you care about. Both are in Search Console already.

• Watch: click-through rate by query, over quarters. A gradual decline on informational terms while impressions hold is the signature of this, and it is visible without any extra tooling.
• Watch: whether commercial queries start showing Overviews. That would be the change that matters, and it has been creeping outward.
• Ignore: daily presence tracking. They flicker. A tool telling you an Overview appeared on 62 percent of your terms this week is measuring noise.
• Ignore: a total traffic number as the headline. It mixes affected and unaffected queries and tells you nothing actionable.

The measurement problem here is milder than with chat assistants, because everything still runs through Search Console. The wider version of that problem, where nothing shows up at all, is covered in tracking brand mentions in AI search.

One final caution about timescale. This feature has changed repeatedly since it appeared, in how often it triggers, how many sources it shows and which queries it covers. Any measurement you take is of the current behaviour, and a decision made on three months of data may be a decision about a version that no longer exists. That argues for responding to the shape of the change, which is stable, rather than to its magnitude, which is not. The shape has been consistent throughout: definitional questions get absorbed, decisions stay with the person making them.