AI Overviews cut clicks in half. Write anyway.

Two findings from the past year tell opposite stories about AI Overviews, and both of them are true.

In March 2025, Pew Research Center recorded what 900 US adults actually did across 68,879 Google searches. When an AI Overview sat at the top of the results, people clicked through to a website in 8% of visits. When no summary appeared, they clicked in 15%. Roughly half the clicks, gone.

A few months later, Ahrefs published its own traffic. AI search sent 0.5% of its visitors. Those visitors produced 12.1% of its signups.

Read together, those two numbers break the habit most people use to judge whether a blog is working. This is what the research actually establishes, where it stops, and what to do about it.

The click math genuinely got worse

The Pew study is the most solid public evidence available, because it measured real browsing behaviour rather than asking people what they remembered doing. Three findings matter.

Clicks to any website roughly halved. 15% of visits without an AI Overview produced a click on a search result. With one, 8% did.

Almost nobody clicks the sources. Links inside the AI Overview itself were clicked in about 1% of visits. If the plan is "get cited and collect the referral traffic", that is the number to sit with. Being cited is not the same as being visited.

People leave sooner. 26% of visits with an AI Overview ended the browsing session, against 16% without one. The answer was good enough, so the session was over.

Two honest limits. It is one month of data from US adults, and Google has changed AI Overviews repeatedly since. Treat the direction as reliable and the exact percentages as a snapshot rather than a constant.

Google says the opposite, and shows no data

On 6 August 2025, Liz Reid, VP and Head of Google Search, published a post stating that "total organic click volume from Google Search to websites has been relatively stable year-over-year", and that "average click quality has increased and we're actually sending slightly more quality clicks to websites than a year ago".

The post contains no figures, no charts, and no year-over-year comparison. Google defines a quality click as one where the user does not quickly return to the results, but publishes nothing to demonstrate either the stability or the improvement.

It is worth being precise about the disagreement, because the two claims are not measuring the same thing.

Pew measured a rate: out of visits to a results page showing an AI Overview, how many produced a click. Google claimed a total: all organic clicks from all of Search, aggregated across every query, year over year.

Both can hold at once if the number of searches grows while the click rate per search falls. Google's own explanation points that way, arguing that people are searching more and asking longer questions. A shrinking slice of a growing pie can be the same size as it was.

What that means for a single blog is less comforting than Google's framing suggests. Aggregate stability across the whole web says nothing about your pages. If the queries you rank for are exactly the informational ones AI Overviews answer outright, you absorb the rate decline without any compensating growth in query volume. Publishers reporting collapses and Google reporting stability can both be describing reality, from different altitudes.

The practical stance: trust the rate finding, treat the aggregate claim as unevidenced until Google shows the data, and measure your own site rather than either.

The visitors who still arrive are different

Now the other half. Ahrefs reported that over 30 days, AI search made up 0.5% of its traffic and 12.1% of its signups, roughly 23 times the conversion rate of traditional organic search.

This deserves more caution than it usually gets. It is a single company's data, not an industry benchmark. It is a B2B software product with a paid signup, which is not most blogs. Ahrefs notes that AI platforms have had problems passing referral data at all, so attribution is imperfect in both directions. In the same analysis they calculate that people arriving from AI search click links 75% less than people arriving from organic search.

So treat 23x as one company's result rather than a law. The useful part is the shape, not the multiple.

There is a plausible reason the shape would hold. Someone arriving from an AI answer has usually had the category explained, the options compared, and your name produced as a recommendation before they ever reach you. That is a later stage of the buying process than a person scanning ten blue links. Fewer people arrive, and the ones who do have already done the work that used to happen on your site.

That is a hypothesis, not a measurement. Hold it loosely and test it against your own numbers.

What actually makes a page get quoted

Most advice on this topic is invented. There is, however, a peer-reviewed answer.

GEO: Generative Engine Optimization, presented at KDD 2024, tested nine ways of rewriting a page and measured how visible each version became inside generative search answers across a benchmark of diverse queries.

Nine rewrites were tested. These are the ones a writer can act on, measured against the paper's main visibility metric.

Two things stand out. Presentation counted, not only substance: writing the same material more fluently moved visibility about as much as some changes to the material itself. And keyword stuffing did not work. The paper groups it among the non-performing methods and notes that simple techniques carried over from classical SEO do not perform well here.

The authors also found that effectiveness varies by domain, so this is a starting point rather than a formula.

Read that list again and notice what it describes. Cite your sources. Quote credible people. Use real numbers. Write clearly. Those are not growth hacks, they are ordinary careful writing, and it turns out to be what generative engines reward. The slop approach is not merely distasteful, it measurably underperforms.

It also explains why thin content fails twice over. A page with nothing specific in it gives a model nothing worth extracting, and gives a reader no reason to stay. Writing in a voice that is actually yours, about things you can substantiate, is the same work as optimising for extraction.

Being cited is not the same as being read correctly

There is a further problem that almost nobody plans for. Getting cited does not guarantee the assistant represents you accurately.

In April 2025, Wu and colleagues published an automated framework for assessing how well LLMs cite sources in Nature Communications. They evaluated seven models across 800 questions and roughly 58,000 statement-source pairs. Between 50% and 90% of responses were not fully supported by the sources they cited, and some were contradicted by them. Even GPT-4o with web search left about 30% of individual statements unsupported, with nearly half of its responses not fully supported.

One important limit: those questions were medical, a domain with unusually strict standards for what counts as supported. Do not read the exact percentages across to marketing content. Read the mechanism, which generalises: a model assembles an answer from fragments, and the fragment it takes from you may not carry the qualification you wrote around it.

That has a concrete writing consequence. A sentence that only means the right thing in the presence of the paragraph above it is a sentence that will be misquoted. Self-contained claims survive extraction. Sentences that carry their own qualifier survive it better still.

This is the real argument for the FAQ format, and it is not a formatting trick. A question with a complete answer underneath it is a unit that stays correct when it is lifted out of context.

What to measure now

If click volume is falling while the value per click rises, session count becomes a misleading headline metric. Some replacements that survive the shift:

• Conversions by landing page, not sessions by landing page. This is the number the change does not corrupt.
• Branded search volume. When an AI recommends you, many people search your name rather than clicking through. That shows up in Search Console as branded queries, not as referral traffic. A rise in branded search with flat organic clicks is often AI visibility working.
• Direct traffic landing on deep pages. A first-time visitor arriving straight onto a niche article usually came from somewhere that stripped the referrer. Treat deep-page direct traffic as a proxy, not as people typing your URL.
• Impressions against clicks in Search Console. If impressions hold or rise while clicks fall, you are being surfaced and not visited. That is the AI Overview pattern, and it is the one case where falling clicks is not a ranking problem.
• Whether you appear at all. Ask the assistants your own category questions on a schedule and record whether you appear and how you are described. Crude, but it is the only direct read on citation, and it takes ten minutes a month.

None of these is as clean as a traffic chart. That is the cost of the change, and no measurement trick removes it. What it does change is the diagnosis: falling clicks with steady impressions is a different problem from falling clicks with falling impressions, and only the second one is about how you are ranking.

Frequently asked questions

Does this mean SEO is dead?

No. The Pew data shows fewer clicks per search, not fewer searches. Ranking well still determines whether you are in the pool of pages an AI Overview draws from, and the GEO research found that the same qualities which earn citations also make a page better for readers. What changed is that ranking now converts into visibility more often than it converts into a visit.

How do I tell whether an AI is citing my site?

There is no reliable dashboard yet. The practical method is to ask the major assistants the questions your customers ask, on a fixed schedule, and log whether you appear and how you are described. Pair that with branded search volume in Search Console, which tends to move when AI recommendations start, and with direct traffic arriving on deep pages.

Should I block AI crawlers?

Only with a clear reason, and only the specific bot you mean. OpenAI documents three separate user agents: GPTBot, which crawls content that may be used to train foundation models; OAI-SearchBot, which surfaces websites in ChatGPT's search features; and ChatGPT-User, which fetches a page when a user's question requires it.

They are controlled independently in robots.txt, and they do different jobs. OpenAI's documentation states that sites opting out of OAI-SearchBot "will not be shown in ChatGPT search answers, though can still appear as navigational links". So blocking GPTBot keeps your content out of model training while leaving you eligible for citation in ChatGPT search. Blocking OAI-SearchBot removes you from the answers themselves.

Most businesses should allow the search crawler. Publishers with licensing leverage sometimes block training deliberately, which is a different decision made for different reasons.

How fast can a new post appear in AI answers?

It varies too much to promise a number. Assistants that search live can surface a page within days of it being indexed. Answers drawn from a model's training data move on a far slower cycle, and a model's training cutoff may be a year or more behind. Anyone quoting a fixed timeline is guessing.

Does publishing more often actually help?

It helps for a specific reason. Generative engines assemble answers question by question, so more published pages means more distinct questions you have a real answer for. Volume alone does nothing, and volume of thin pages is actively counterproductive given what the GEO research found about keyword stuffing. Volume of genuinely useful, well-sourced pages compounds, which is the premise the whole approach rests on. The hard part was never knowing this. It is sustaining the cadence.

Do AI Overviews affect every query equally?

No, and this is why site-wide averages mislead. AI Overviews appear most on informational queries, which is exactly where thin explainer content used to earn easy traffic. Transactional and navigational queries are affected far less. A site whose traffic is mostly informational absorbs the full rate decline, while a site ranking for commercial-intent terms may barely notice it. Segment by query intent before concluding anything about your own numbers.

The short version

Google's AI Overviews roughly halved click-through on the searches where they appear, and almost nobody clicks the citations inside them. Google says aggregate click volume is stable but has published no data, and the two claims measure different things. The visitors who still arrive appear to be further along and worth more, though the strongest public number for that is one company's. The peer-reviewed research on what generative engines reward points at citations, quotations, statistics and clear writing, and explicitly not at keyword stuffing. And being cited is not the same as being quoted correctly, which is an argument for self-contained claims.

The practical conclusion is uncomfortable but simple. Your traffic chart is going to look worse, and that is not the same as your blog doing worse. Change the metric you judge it by before you change the strategy.

Then keep publishing pages worth quoting.