AI Overviews Cut CTR 23.1% in France: What AEO Teams Do Next

12 October 2026 · The Cresia team

Click-through can fall fast, within days, once AI Overviews arrive in a market, and the loss lands hardest on the pages the feature covers. Ahrefs reported that the most exposed of 963 French domains lost 23.1% of their CTR after Google launched AI Overviews in France on July 22, 2026. For AEO and GEO teams the job is to stop reading rankings and traffic as one number and track them separately.

Key takeaways

Ahrefs reported that the most exposed of 963 French domains lost 23.1% of their CTR after AI Overviews launched on July 22, 2026. That is the worst-hit group, not the average site.
A CTR drop is not a ranking drop. Impressions can hold steady while clicks fall, so track position, impressions, clicks and citations as separate series.
AI answer engines favour passages that state a complete answer plainly, so rewrite key pages around one question per section with the answer first.
Segment Search Console by query type and page template before reacting; informational queries are where the exposure sits.
Do not chase tricks or rebuild a working site. Fix measurement first, then the pages that carry commercial value.
Treat nine days of post-launch data as an early read and re-check at 60 and 90 days.

What Ahrefs measured in France

Ahrefs pulled 963 domains from Google Search Console and compared 28 days before the July 22 launch with the 9 days after. The 23.1% CTR loss belongs to the most exposed domains in that set. It is not a market-wide average, and quoting it as one will mislead your stakeholders.

The design has two features worth holding on to. It uses Search Console, which is the right source for this question because the change happened inside Google's own results page. And the after-window is only nine days long.

Nine days is enough to see a direction. It is not enough to see a settled level. Weekday mix, news cycles and late-summer holiday behaviour in France can all sit inside a window that short. Users also change habits slowly: some will start scrolling past the AI answer, others will lean on it more. The figure may move in either direction as the feature expands to more query types.

So the sensible reading is this: a major search market just got AI Overviews, and the sites with the most overlap with the feature saw a double-digit fall in click-through almost immediately. That is a strong enough signal to act on measurement now and to hold back on large rebuilds until more data arrives.

Why a CTR drop is not a ranking drop

A page that ranks third for a question and sits below an AI-generated answer is still ranked third. Its impressions may even be intact. What changed is the share of searchers who click, because some of them got what they needed from the answer above it.

That distinction matters for how you report. If your weekly deck has one line called organic performance, a CTR fall will read as a ranking problem and someone will start rewriting title tags. Meanwhile the underlying cause is a layout change.

Consider a mid-size software company with a glossary of 400 terms. Each glossary page ranks well for a definition query and used to collect a steady trickle of clicks. After an AI answer appears for those queries, the page keeps its position and its impressions, and the clicks thin out. Nothing on the page is broken. The page's job has changed from delivering a visit to supplying a passage that others quote.

The same company's pricing comparison pages behave differently. Searchers there want to evaluate, compare and decide, and an AI summary is a poor substitute for the table they came for. Expect uneven exposure inside a single domain. The domains at the sharp end in the Ahrefs data are most likely the ones whose content is mostly definitions and short factual answers, though you should confirm that against your own segments rather than assume it.

How AI answer engines choose what to cite

The details differ by product and change often, so check each vendor's own documentation before building a process around a mechanism. The broad pattern is stable enough to plan around.

An answer engine takes a question, retrieves candidate passages from an index, and composes a response, often with a handful of cited sources. Several things tend to separate a passage that gets used from one that doesn't:

The answer is stated outright. A paragraph that opens with the answer and then supports it is easier to lift than one that builds to a conclusion over six sentences.
The passage stands alone. If a sentence only makes sense after reading the three before it, it is hard to quote. Name the subject in the sentence itself.
The entity is unambiguous. A page that says clearly who wrote it, what organisation stands behind it and what product or concept it is about gives the system less to guess at.
The page is crawlable and indexable. This sounds basic, and it is the most common silent failure. Blocked resources, client-rendered content that never reaches the crawler and canonical mistakes all remove you from the candidate pool.
The claim is specific and checkable. Vague superlatives give a system nothing to cite. A named figure with an attributed source does.

Google's guidance for site owners, published on its Search Central blog and documentation, points to the same fundamentals that already applied to classic search. There is no separate markup that unlocks inclusion. Read the current wording there instead of relying on secondhand summaries, because the guidance has been revised more than once.

Being cited and being clicked are also different outcomes. A citation builds familiarity and can lead to a later branded search or direct visit, but it rarely produces a one-to-one replacement for the lost click. Plan for that gap in your forecasts.

What to change in practice

Start with the pages that matter commercially, not the ones that lost the most clicks. A glossary page that lost 40% of a small number of clicks is a lower priority than a product-adjacent guide that lost 10% of a large one.

Sort your pages by intent. Mark each important URL as definitional, comparative, procedural or transactional. Exposure to AI answers sits mostly in the first and third.
Rewrite openings for definitional and procedural pages. Put a two-to-three sentence answer directly under the heading, then the nuance. This helps readers who do click and gives answer engines a clean passage.
Give every H2 one job. Phrase it the way a person would ask the question and answer that question completely beneath it. Sections that wander are hard to quote.
Add what a summary can't supply. A worked example, a decision table, a checklist with owners, a template to download. These are the reasons to click through after the answer has been read.
Attribute figures in the prose. Name the outlet that reported a number. Unsourced statistics are a liability for readers and give answer engines little reason to trust you.
Fix crawl and render basics. Confirm key content is in the HTML the crawler receives, canonicals are right, and nothing important sits behind a script that fails.
Strengthen entity pages. Authors, the organisation and the products should each have a clear, consistent page that the rest of the site links to.

One caution on step two. Writing an answer-first opening does not mean thinning the page. The goal is a short, quotable passage at the top of a page that still has depth underneath. If the page was only ever a short answer, you may need to decide whether it deserves to exist or should be merged into a stronger one.

Teams running structured experiments on page layouts can test openings and content blocks rather than argue about them; A/B testing covers the method, and Evolve is where Cresia's experimentation product lives.

How to measure AEO and GEO without fooling yourself

The worst measurement habit after a launch like this is to compare one blended organic number before and after. Split it. The table below is a workable starting plan, and each row answers one question.

Question	Metric	Where it lives	Suggested owner
Did we lose rank?	Average position by query group	Search Console	SEO lead
Did we lose visibility?	Impressions by page template	Search Console	SEO lead
Did we lose clicks only?	CTR by query intent	Search Console, segmented	Analytics lead
Are we being cited?	Citation presence for a fixed query set	Manual sampling or an AI visibility tool	Content lead
Did lost clicks cost us leads?	Conversions from organic landing pages	Web analytics	Analytics lead
Is branded demand moving?	Branded query impressions and direct visits	Search Console, analytics	Growth lead

A few notes on running it.

Fix your query set before the next change, not after. Pick 50 to 150 queries that represent your market, group them by intent, and keep them stable for at least a quarter. Citation tracking on a moving set of queries produces noise.

Keep a dated change log. When a template is rewritten, an AI Overviews rollout reaches a new country, or a core update lands, write it down. Without that log, you will be arguing about causes from memory three months later.

Be careful with attribution, too. If organic conversions are flat while clicks fall, the remaining clickers may simply be more qualified. That is a plausible outcome, not a guaranteed one, and your own data has to show it. A clean tracking specification makes this sort of comparison trustworthy, because you know the events meant the same thing before and after.

For teams that want AI-search visibility in the same view as classic search reporting, MediaPilot is the Cresia product to look at.

What not to do

A 23% figure produces urgency, and urgency produces bad decisions. A short list of things to avoid:

Don't publish a wall of FAQ pages to farm citations. Thin question-and-answer pages are cheap to make and equally cheap to ignore. They also dilute a site that has genuine depth.
Don't rewrite everything at once. You will lose the ability to tell what worked. Change one template group, watch it for several weeks, then move on.
Don't treat one country's number as your forecast. The French result describes one market in its first nine days. Your own exposure depends on your query mix, your language and how far the feature has reached there.
Don't abandon classic SEO. The pages that get cited are, for the most part, pages that already rank and are already crawlable. Weak foundations hurt on both fronts.
Don't report citations as a win without clicks and conversions beside them. Being named in an answer is useful. It is not revenue.
Don't write for the machine in a voice no person would read. Stilted, over-structured copy damages the visit that does arrive.

A 30-day plan for a small team

The plan below assumes one SEO lead, one analyst and a part-time content owner. It is deliberately modest.

Week one: baseline. Export 90 days of Search Console data by page and query. Tag pages by intent and template. Build the fixed query set and record current citation presence.

Week two: diagnose. Compare CTR against impressions and position for each template group. Identify which groups lost clicks while holding position. Check crawl, render and canonical health on the top 50 pages.

Week three: change. Pick one or two high-value template groups. Rewrite openings, tighten H2s, add one non-summarisable asset per page, and log the date.

Week four: read and decide. Look at the change against a comparison group of untouched pages. Decide what to extend, what to leave and what to merge. Schedule a re-check at 60 and 90 days.

Cross-functional teams can pick up further context in the growth teams solution page, which is relevant where SEO, content and analytics share ownership of this problem.

The broader point is plain. Search is becoming a place where some questions are answered before a click happens. Teams that know which of their pages are exposed, which are protected by intent, and what each one is worth will make better calls than teams reacting to a single scary percentage.

Frequently asked questions

Did AI Overviews cut CTR by 23.1% for every site in France?

No. Ahrefs reported that figure for the most exposed domains among the 963 it tracked in Search Console. Other domains in the sample were affected less, and your own number depends on how many of your queries trigger an AI answer.

Should we change our content strategy because of one nine-day window?

Change your measurement now and your content selectively. Nine days is long enough to show direction and too short to show a settled level. Start with answer-first rewrites on high-value pages and re-check at 60 and 90 days before making larger commitments.

What is the difference between AEO and GEO?

The terms overlap and are used loosely. AEO usually refers to earning a place in direct answers, such as featured snippets and AI Overviews. GEO usually refers to being cited or mentioned by generative AI products more broadly. The page-level work is largely the same.

How do we know if an AI answer engine cites us?

Sample a fixed set of queries on a regular schedule and record whether your domain is cited, in which position and for which passage. You can do this by hand for a small set or use a visibility tool for a larger one. Check each vendor's documentation for how citations are displayed, since the format changes.

Sources

https://ahrefs.com/blog/ai-overviews-france-impact/
https://developers.google.com/search/blog
https://www.conductor.com/academy/aeo-geo-benchmarks-report/