AI Overviews Click Loss: Find Your Real Number

Start with the spread. Ahrefs' 2025 analysis reported an average clickthrough rate roughly 34.5 per cent lower for queries showing an AI Overview. Research published in 2025 by academics at Carnegie Mellon University and the Indian School of Business found approximately 40 per cent fewer outbound clicks on searches where an AI Overview was present. Seer Interactive's 2025 study of AI Overview queries reported organic clickthrough rates around 61 per cent lower. Publisher-side figures circulated during the same period showed a clear device split, at 47.5 per cent on desktop against 37.7 per cent on mobile; anyone intending to quote those last two numbers should trace them back to the original publisher submission first, because they travel further than their provenance does.

The variance is not evidence that someone is wrong. It is evidence that four studies measured four different populations. A publisher sample is dominated by informational and news queries, which is precisely the query class an AI Overview answers most completely, so it will always show a steeper decline than a B2B software site whose demand sits on branded, commercial and comparison terms. Device split matters because the desktop layout pushes organic results further down the page than mobile scrolling behaviour suggests. Position within the AI Overview source list matters, and so does whether the study reported clickthrough rate or absolute clicks, because impressions frequently rose over the same period, which flatters or flattens the headline depending on the denominator chosen.

One counter-finding deserves attention. Several 2025 analyses observed that the engagement quality of the clicks that did still arrive looked broadly unchanged, with bounce rate, time on site and returns to the search results holding steady. That distinction matters because it points to a filtering effect rather than a degradation: the visitors who wanted a one-line answer took it from the Overview, while those with a genuine task still clicked through. Traffic fell without the remaining traffic getting worse. For a commercial site, that is a materially different problem from a content quality failure, and it should be reported so.

Applying any of these percentages to a specific site is one of the easiest mistakes to make. The moment a figure such as 61 per cent enters a board paper, someone will ask which queries it covers, what proportion of the site's demand triggers an AI Overview at all, and why the same slide does not show the site's own measurement. The safer conclusion is that the published range establishes that the effect is real and material, and nothing more.

The signal pattern is consistent and reasonably distinctive. Impressions stay flat or rise, average position holds within a point or so, and clickthrough rate falls on a defined set of pages. Google Search Console's own documentation on the Search performance report makes clear that impressions and clicks from AI experiences are counted inside the standard Web search type totals rather than isolated as a separate surface, and at the time of writing there is no filter that separates them. Any number produced for a single site is so an inference assembled from clickthrough rate, impression and position movement. Stating that limitation out loud, in the same slide as the estimate, is what separates a defensible diagnosis from a guess.

Segment before comparing anything. Split brand from non-brand, because branded demand rarely triggers an Overview and will mask the effect if left in the total. Separate question and how-to phrasing from transactional phrasing. Split desktop from mobile. Then compare two like-for-like windows of equal length, ideally two consecutive 90-day periods rather than year-on-year totals, which import seasonality, unrelated algorithm updates and site changes into the comparison.

As an illustration of what a clean pattern looks like: a non-brand informational page set records impressions up 12 per cent across two comparable 90-day windows, average position unchanged, and clickthrough rate falling from 3.1 per cent to 1.9 per cent, with the loss concentrated on question-format queries and disproportionately on desktop. That combination is difficult to explain through ranking decline, because the rankings did not decline. It is consistent with an answer being served above the results. It is not proof, and the write-up should say so.

Third-party SERP feature tracking closes part of the gap by establishing which of your target queries actually trigger an AI Overview and how volatile that trigger rate is from week to week. Volatility here is higher than most teams expect, so a single snapshot is close to useless; a rolling weekly record across the same query set is what makes the trigger rate quotable. A free AI visibility scan is a reasonable way to see whether a brand appears in AI answers at all before committing budget to a fuller measurement programme.

AI Overviews have become the default explanation for any organic decline, which is convenient and frequently wrong. Core and spam updates, seasonality, keyword cannibalisation after a content refresh, a site redesign that changed internal linking, consent banner or analytics tagging changes, and reporting quirks introduced when Google adds a new surface to its data all produce declines that look superficially similar in a topline chart.

Work through an elimination sequence before attributing anything:

Check the drop date against confirmed Google update windows. A decline that begins on an update rollout date and affects rankings as well as clickthrough rate is an update, not an Overview.

Confirm average position is genuinely stable on the affected query set. Falling position with falling clicks is a ranking problem.

Compare the same period in the previous two years for seasonality, particularly for education, travel, finance and retail cycles.

Verify that analytics sessions and Search Console clicks moved together. Divergence points at tracking or consent, not at Search.

Check whether the affected URLs changed, were redirected, consolidated or lost internal links during the window.

Misattribution has a cost beyond tidiness. A team that files a thin, outdated page under "lost to AI Overviews"stops working on the fix that would have recovered the traffic. Document the diagnosis in a form that survives scrutiny: the query set and its definition, the two date ranges, the observed impression, position and clickthrough movement, the confounders tested and excluded, the trigger rate evidence, and an explicit statement of what the data cannot demonstrate. Finance audiences respond considerably better to a bounded estimate with stated limits than to a confident percentage with none.

Exposure is not evenly distributed across a site. Definitions, how-to instructions, comparison explainers, statistics round-ups and long-tail informational content are the most vulnerable, because the answer fits in a paragraph and the searcher has no reason to continue. Branded and navigational queries hold up, since the user already intends to reach a specific destination. Transactional queries, high-consideration research where the buyer wants to verify a source directly, and locally intent-driven searches in places such as Manchester, Leeds or Bristol all show more resilience, though local behaviour is shifting as assistants improve at interpreting location.

Triage at page level rather than at site level. Separate the pages that lost traffic but never converted from the pages that lost traffic and revenue. A glossary entry shedding 4,000 sessions a quarter with no assisted conversions is a reporting problem; a comparison page shedding 300 sessions that fed the demo pipeline is a commercial one. Rebuilding reporting around conversion-weighted traffic, where each page's loss is expressed against the pipeline value it historically influenced, tends to shrink the apparent crisis and sharpen the actual priority list. Boards find that framing easier to act on than a sessions chart.

Zero-click behaviour on simple informational queries is structural. Featured snippets started the pattern years ago and generative answers have extended it, so planning to recover every lost click sets a target that cannot be met and makes an otherwise successful programme look like a failure. Accepting that is not defeatism; it is what allows the measurable objective to move somewhere useful.

Research into AI Overview source selection indicates that close to 30 per cent of cited URLs do not come from page-one organic results, and the proportion varies considerably by query set and vertical. A page sitting at position eight can still be the source an AI answer quotes, which inverts a familiar prioritisation. Source suitability, clean structure, a directly stated answer near the top, current data and evident subject authority can matter more than one further rank position. Traditional rank reporting hides this entirely, because it measures a ladder the answer engine is not obliged to climb in order.

Category and entity accuracy deserves separate attention, because errors there quietly cause more damage than click loss itself. If an AI system has a business filed under the wrong category, associated with a service it discontinued, or conflated with a similarly named company, it is excluded from the relevant answers before clickthrough rate becomes relevant. Correcting a misclassification is a discrete, verifiable fix with an observable before and after, which is more than can be said for most visibility work.

Google is also only one surface. AI assistant usage is fragmenting across ChatGPT, Gemini, Perplexity, Claude and Meta AI, each drawing on different sources and each forming its own view of a brand, so a response plan built solely around Google AI Overviews leaves the majority of AI-influenced buying research unmeasured. Visibility inside one system is not visibility inside another. Prompt-set testing across platforms is the practical way to see the difference, and it is the basis of AI Recommendation Optimisation as a discipline distinct from conventional search work.

Weeks 1 to 3: baseline. Define the affected page set precisely and record its impression, position and clickthrough figures across two comparable windows. Establish the AI Overview trigger rate for the associated queries and track it weekly rather than once. Build a prioritised prompt set of real buyer questions, typically between 15 and 100 depending on portfolio breadth, phrased as customers actually phrase them ("which AI visibility tools work for UK B2B firms", "is prompt monitoring worth it for a small marketing team"). Run that set across the six surfaces that matter, recording whether the brand appears, which competitor appears in its place, and which third-party sources each answer cites.

Weeks 4 to 8: fix the diagnosable gaps. Work through crawler access for AI user agents in robots.txt, structured answers placed where a retrieval system will find them, accurate entity and category descriptions across owned and third-party properties, consistency of core facts between the website, directories and profiles, freshness on anything carrying a date or a figure, and presence in the independent sources those AI answers were observed to cite. Most of this is unglamorous housekeeping, and most of it is where the recoverable ground sits. A structured Advanced Audit covers the same territory for teams without the internal capacity to run it.

Weeks 9 to 12: retest and report. Re-run the identical prompt set under the same conditions, compare citation share and answer inclusion against the named competitors from the baseline, and rebuild reporting around assisted demand rather than clicks alone. Movement over a single quarter may be modest and will not be uniform across platforms; report it that way. Ongoing monitoring, priced openly on the AI visibility plans page, is what turns a one-off retest into a trend line.

Governance holds the whole thing together. Structuring the programme around the four functions of the NIST AI Risk Management Framework, namely Govern, Map, Measure and Manage, forces an explicit statement of what the measurement can and cannot evidence: prompt-set results are repeatable observations, not audited market share, and AI answers vary by user context. Say that in the reporting pack before someone else says it in the meeting. Teams weighing up how to resource this can book a Growth consultation to work through scope against their own query mix.

How much click loss do AI Overviews actually cause?

Published 2025 estimates range from roughly 34.5 per cent lower clickthrough rate in Ahrefs' analysis to around 61 per cent in Seer Interactive's study of AI Overview queries, with Carnegie Mellon and Indian School of Business research reporting approximately 40 per cent fewer outbound clicks. The spread reflects different query mixes, devices, sample types and measurement windows rather than disagreement about direction. No single figure transfers safely to an individual site.

Can Search Console show AI Overviews click loss separately?

No. Google reports impressions and clicks from AI Overviews and AI Mode inside the standard Web search type totals, so no Search Console report isolates them. Any site-level estimate is an inference drawn from clickthrough rate falling while impressions and average position hold steady, and it should be presented with that limitation attached.

How do I tell AI Overviews click loss apart from a Google update hit?

Check average position first. An update typically moves rankings, so clicks and position fall together, whereas AI Overview pressure usually shows stable or improving position with falling clickthrough rate. Cross-reference the start date against confirmed update rollouts and against week-by-week AI Overview trigger data for the affected queries.

Which types of pages lose the most clicks to AI Overviews?

Definitions, how-to guides, comparison explainers, statistics round-ups and long-tail informational content are the most exposed, because the answer can be summarised completely above the results. Branded, navigational, transactional and high-consideration queries hold up considerably better. Triage by conversion value rather than session volume, since a large share of the lost traffic never converted.

Is it possible to recover traffic lost to AI Overviews?

Part of it is structurally gone, and planning for full recovery sets a target that cannot be met. The recoverable share comes from becoming the source AI answers cite, which is realistic given that research into AI Overview source selection suggests close to 30 per cent of cited URLs sit outside page-one results. The practical objective for AI Overviews click loss so shifts from restoring sessions to increasing citation and recommendation share across Google, ChatGPT, Gemini, Perplexity, Claude and Meta AI.