What AI Overviews cost you in clicks, according to the people who measured it

Three credible groups measured what AI Overviews do to clicks. One got roughly 15 percent, one found clicks close to halved, one got 38 percent.

Every agency quoting one of those numbers at you has picked the one that suits the pitch. The gap between them is more useful than any single figure, and it tells you something specific about your own site.

1. The three numbers, with their methods attached

A study of 161,382 matched article-language pairs found that AI Overview exposure reduced daily traffic to English articles by approximately 15 percent. It compared articles that got an AI Overview against matched ones that did not.

Pew Research found users clicked a search result on 8 percent of visits that produced an AI summary, against 15 percent of visits without one. That is close to a halving of an already small number.

A field study reported by Search Engine Journal found a 38 percent drop in organic click-through on tracked keywords.

None of these is a bad study. They are answering three different questions.

2. Why the spread exists, and why it is the finding

The 161,382-pair study asks what happened to traffic at a set of articles. Pew asks what a human did during a browsing session. The field study asks what happened to click-through on a keyword list somebody chose.

Different denominators produce different percentages from the same underlying reality.

So the correct thing to say to a client is a range with the methods named, and we say it that way at NovaTechRay even when a single confident number would close faster. A vendor who quotes 38 percent with no method attached is not better informed than one who quotes 15 percent. They are just more alarming.

Here is the complication though. We spent a quarter quoting the 15 percent figure as the conservative one, and that was its own kind of error.

3. The conservative number is not the safe number

Fifteen percent sounds survivable, so we used it as the floor. Then we looked at what it was averaging.

The same study found the largest relative declines in Culture articles and substantially smaller effects in STEM. That is a huge internal range folded into one mean.

If your site is mostly short factual answers, the average is not your number. The mean was dragged up by content categories that barely moved.

We had been handing clients a floor that was really a midpoint of a distribution most of them sat on the wrong side of.

4. Work out your own exposure instead

The studies cannot tell you your number. Your query mix can.

Pull your top pages by impressions in Search Console. For each one, write the question the page answers, then ask whether a four-sentence paragraph would finish the job for the reader.

If the honest answer is yes, that page is exposed. A definition page, a units conversion, a "what is X" explainer, an opening-hours query. These are the ones that go.

If the answer is no, because the reader needs your prices, your availability, a tool, a download, a comparison against their specific situation, or to contact a human, that page holds up much better.

Count the two piles by impressions. That ratio is a better forecast for your site than any published mean.

5. What to do with the exposed pile

Do not delete it, and do not rewrite it into longer filler.

Those pages still earn the impression and the citation, and impressions are the only thing Google's Search Console Generative AI report gives you. The exposure is real and it is worth having.

What changes is what you expect from them. Stop measuring a definition page on sessions. Measure it on impressions and on whether your name appears in the answer.

Then move the conversion work onto the pages in the second pile, which is where the clicks still land.

6. The part nobody wants to hear about recovery

Getting cited inside the AI Overview does not give the traffic back.

Pew Research found only 1 percent of users clicked a link inside an AI summary. One percent.

So the standard agency pitch, that GEO recovers the traffic AI search took by getting you into the answer box, is arithmetic that does not work. The citation is brand exposure. It is worth buying at the right price. It is not a click.

We sell it as exposure, which is a harder sell and the only one we can defend.

7. The reporting change this forces

If a chunk of your visibility now happens without a click, a sessions-only dashboard will show a business in decline that is not declining.

Report impressions and clicks side by side, from the same date range, and never let one stand in for the other. Add a column for whether your brand was named in the answer, sampled properly rather than once, on the method we set out in why every AI visibility tool gives you a different number.

Then say out loud, in the report, which pages you expect to lose clicks and keep impressions. Predicting the decline before it happens is the difference between a measurement and an excuse.

The short version

• Treat the click loss as a range of 15 to 38 percent, and name the method whenever you quote one.
• Do not use 15 percent as your floor. It is a mean across categories that moved very differently.
• Sort your top pages into "a paragraph finishes this" and "the reader still needs something from us".
• Expect the first pile to lose clicks and keep impressions, and change what you measure it on.
• Stop selling citation as traffic recovery, because only about 1 percent of users click inside an AI summary.
• Report impressions and clicks side by side, and forecast the decline before it lands.

Measurement that survives a client asking "compared to what" is most of what we do at novatechray.com.