AI Overviews and Your Traffic: Why the Studies Range From 15 to 89 Percent, and Which Figure Applies to You

Depending on which article you read, AI Overviews cost you 15 percent of your clicks, or 89 percent. Both numbers come from real measurements, and neither is wrong. The spread does not mean the field is confused. It means the studies asked different questions, and the answer that matters for your company depends on which of those questions resembles your situation.

What the two most solid measurements actually did

Two studies stand out because their method is stated plainly and can be checked. They are worth reading before any percentage is quoted.

Notice what separates them. Pew watched people and counted visits. Ahrefs watched keywords and counted rates. Both are legitimate; they simply do not produce the same number, and averaging them would produce a figure that describes nothing.

One detail from the Pew data deserves its own line, because it is the one most often left out: sessions ended entirely on 26 percent of pages with an AI summary, against 16 percent without. The summary does not only redirect the click. It frequently ends the search.

Why the published range is so wide

Three variables move the result, and almost every headline figure hides at least one of them.

This is why a figure like 89 percent can be accurate and useless at the same time. It describes a specific slice, usually navigational queries where a summary answers the whole question. Quoting it as the effect on your website is not a small imprecision; it is a different measurement entirely.

Which part of a B2B website is actually exposed

The useful split is not by industry but by what a page is for. An AI summary is good at reproducing an explanation and poor at replacing a decision.

Exposed are the pages that explain: definitions, background articles, guides, the piece that answers what a thing is and how it works. That is exactly the material a seventy-word summary handles well, and it is often the part of a mid-market website that carries the most impressions.

Less exposed are the pages where a decision forms: your services, your prices, your references, your process, anything specific to you. A general answer does not substitute for them, because the question behind them is not general. This is also where enquiries come from, which softens the picture considerably compared to the headline numbers.

The uncomfortable consequence for content strategy: the article written purely to attract traffic through a general question is the most exposed asset you own. The article that answers a specific question about a specific decision is the most durable one.

The other direction, with its own caveat

The same Pew data contains the counter-movement. 88 percent of the summaries examined cited three or more sources, and their average length was 67 words. Being one of those sources means appearing in a position that did not exist before, above the classic results, in the place people actually read.

The caveat sits in the same dataset and is regularly omitted from the optimistic version: a link inside the summary was clicked in 1 percent of all visits. Being cited therefore produces visibility, name recognition and the chance to be the supplier a buyer already recognises when they eventually search directly. It does not reliably produce sessions. Anyone selling citation as a traffic channel is stretching a number that will not hold.

What the GEO paper shows, and what the market makes of it

One study is quoted in nearly every article on this subject, usually with a precise-sounding figure attached to a single tactic. It is worth reading what it says.

GEO: Generative Engine Optimization by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande was submitted in November 2023, revised in June 2024, and published at the KDD 2024 conference. It demonstrates in a controlled setup that content can be deliberately optimised for higher visibility in generative answers, reporting an improvement of up to 40 percent, and it introduces a purpose-built benchmark for that purpose.

That is a genuine and useful result: it is the reason to believe that optimising for AI answers is possible at all rather than wishful thinking. But it is a laboratory result on a benchmark the authors built, not a measurement against Google in 2026. The precise single figures circulating for individual tactics do not appear in that form in the abstract. We mention this not to diminish the paper but because a study quoted this often deserves to be quoted correctly, and because a proposal that cites it with a decimal point is telling you something about its own sourcing.

How to measure it for yourself in an afternoon

No industry figure answers the question for your website. Your own Search Console does, with one separation most people skip.

• Split your pages by purpose. Guides and explanatory articles in one group, services, prices and references in the other. Two lists, once.
• Compare impressions and clicks separately for each group over the last twelve months. Not the combined total, which hides the effect.
• Read the pattern. Impressions stable while clicks fall means the summary is taking the click. Both falling means a ranking or competition issue that has nothing to do with AI. Impressions rising while clicks stay flat is the pattern we have seen most often in the mid-market.
• Check where you already are a source. Put a handful of your own core questions to an AI assistant and note whether you appear and whether what it says about you is accurate. That takes twenty minutes and tells you more than a report.

If that check produces a clear signal, it also tells you where budget should go. If it produces noise, that is itself an answer: the effect on your site is smaller than the coverage suggests, and the money is better spent elsewhere.

Conclusion

The measurements are real and they diverge for good reasons. Pew watched people and found clicks on traditional results falling from 15 to 8 percent when a summary appears. Ahrefs watched keywords and found position one down about 34.5 percent. Both describe the same shift from different angles, and neither is the number for your company. What is exposed on a B2B site is the explaining half, not the selling half, and the same data that shows the loss also shows 88 percent of summaries citing three or more sources. Treat the range as a reason to measure rather than a reason to panic, and be sceptical of any proposal that answers this question with a single decimal figure. If you want that measurement done against your own pages, that is what a first conversation is for, and the budget side is in what GEO costs.

Frequently asked questions about click loss from AI Overviews