AI Overview Click Loss: 104,008 Impressions, 395 Clicks

Between July 1 and July 28, 2026, seven pages on this site pulled 104,008 impressions from Google at an average position of 6.6. They returned 395 clicks. That is a click-through rate of 0.38 percent. Over the same 28 days, eight other pages on the same site sat at an average position of 9.2, which is worse, and returned 1.79 percent. The better-ranking group earned 4.7 times fewer clicks per impression than the worse-ranking group.

Every study you have read about AI Overviews reports something like this. Almost none of them publish the underlying numbers, because most of the people writing about it are agencies who cannot share client data. We are a two-product company with one Search Console property, so we can just show you.

What follows is the whole dataset, the live SERP checks we ran against it, and the part that surprised us. The AI Overview turned out not to be the variable we thought it was.

What is AI overview click loss?

AI overview click loss is the gap between the clicks a page actually earns and the clicks its ranking position would historically have predicted, caused by Google resolving the query inside an AI Overview before the user reaches any result.

It is not the same thing as losing rankings. In our data the affected pages hold their positions perfectly well. Some of them rank in the top five. The impressions keep arriving. What stops arriving is the visit, because the question got answered on the results page and there was nothing left to click for.

That distinction matters because it breaks the reporting most people rely on. A rank tracker will tell you everything is fine. Position 6 looks like position 6 whether it earns 6 percent or 0.4 percent.

The data: 118,408 impressions across 15 pages

All figures are from Google Search Console for contextbolt.com, July 1 to July 28, 2026, web search, all countries and devices. No sampling, no modeling, no estimates. This is the raw export.

We split the pages by the shape of the question they answer, not by their performance. Group A answers questions with one short correct answer. Group B answers questions needing a procedure or a comparison. The split was made from the page topics before looking at the click column.

Group A. Questions with one short answer.

Group B. Questions needing a procedure or a shortlist.

Group A ranks 2.6 positions better and converts 4.7 times worse. If those 104,008 impressions had performed at Group B’s rate, they would have returned 1,864 clicks. They returned 395. That is 1,469 missing visits in 28 days, and 79 percent fewer clicks than our own site earns at a worse position.

The single starkest page is are X bookmarks private. It is one of our best-ranked pages, sits at position 7.7, and pulled 25,023 impressions in four weeks. It got 37 clicks. Thirty-seven.

Why do our best-ranking pages earn the fewest clicks?

Because they answer questions that fit in a sentence, and Google now writes that sentence itself.

“Are X bookmarks private?” has one correct answer and it is one word. “What is the Reddit saved posts limit?” has one correct answer and it is a number. “Where are my bookmarks on X?” has one correct answer and it is a menu path. There is no version of those pages that gives the reader a reason to click once they have read the top of the results page, because there is nothing left to tell them.

“How do I export my Reddit saved posts?” is a different animal. The honest answer is three methods with different trade-offs, one of which involves a script and a GDPR data request. A summary can gesture at that. It cannot replace it. So the click survives.

This is the uncomfortable part for anyone selling SEO by ranking position. We did not get worse at ranking. We got better. Our best pages, on our strongest topics, at our best positions, are the ones bleeding.

Is the AI Overview really the cause?

This is where we expected a clean answer and did not get one.

On July 31, 2026 we pulled live United States SERPs for four queries in the dataset, two from each group, and recorded whether an AI Overview was present.

The one query with no AI Overview has the best click-through rate on a large impression base. That fits.

But three of the four carry an overview, and their rates still range from 0.19 percent to 1.24 percent. That is a six-fold spread among pages that all have an AI Overview sitting above them. If presence were the mechanism, that spread should not exist.

So the binary framing everyone uses, does this keyword have an AI Overview or not, is too blunt to be useful. It tells you an overview is there. It does not tell you whether the overview took your click.

What is answer compression?

Answer compression is how completely a query’s correct answer fits inside a short generated summary. High compression means the summary can be the whole answer. Low compression means it can only be a preview.

Read the table again through that lens and it resolves. “Are twitter bookmarks public” compresses to one word, so it loses 99 out of every 100 impressions. “Best twitter bookmark manager” compresses to a list of names, which is useful but not sufficient, because you still want to know which one exports your data, so it keeps six times more of its clicks. “Export reddit saved posts” barely compresses at all, and Google did not even try.

That last point is the tell. Google appears to decide whether to generate an overview partly on the same basis. The overview is downstream of answer compression, not the cause of it. Which means the presence check most AEO tools sell you is measuring the symptom.

The practical consequence is that you can predict this before you write a page. Ask whether a competent person could answer the query completely in three sentences. If yes, that page will not earn clicks in 2026 no matter how well it ranks. Write it for citation and brand presence, or do not write it.

How does this compare to the published studies?

Our gap is larger than the published averages, and we think we know why.

Pew found a 47 percent reduction. Seer found 58 percent. Ahrefs found 58 percent at position one. We measured 79 percent.

We are not claiming our site is special. Our Group A is simply an unusually pure sample of high-compression queries, because our biggest pages happen to answer yes-or-no and how-many questions about social platforms. A site whose top pages are recipes or product comparisons would land closer to the published averages. That is the point. Site-level averages hide a variable that page-level shape exposes.

Worth noting alongside this, Seer also found the floor is moving for everyone. Queries without any AI Overview still fell to 1.45 percent, down 46 percent year on year. There is no safe category, only a less exposed one.

Why won’t Search Console tell you this?

Google shipped a generative AI performance report in June 2026. It was the first time site owners could isolate AI surface visibility at all, and it was genuinely welcome.

It reports impressions. That is the entire list. Per Google’s own documentation, the report includes no clicks, no position, no click-through rate and no query data. Meanwhile AI feature traffic stays folded into the Web totals in the standard performance report, so you cannot subtract one from the other either.

Reporting the impressions but not the clicks is a decision, and it is the one metric pairing that would let every publisher calculate exactly what we calculated above. Google also states in its AI features documentation that an AI Overview occupies a single position, and every link inside it is assigned that same position. So your reported average position now blends two different things that behave nothing alike.

None of this makes the data useless. It makes it inferential. You cannot read AI overview click loss off a dashboard. You have to reconstruct it, which is what the next section is for.

How do you measure this on your own site?

The method took about twenty minutes and needs no paid tool beyond a SERP check.

1. Export 28 days of page-level data from Search Console. Clicks, impressions, CTR, position. Use a full month so weekday patterns wash out.
2. Drop anything under 300 impressions. Below that a single click swings CTR by whole percentage points. Two of our early candidates looked like huge outliers until we noticed they had 29 impressions between them.
3. Label each page by answer shape, not by performance. One short answer, or a procedure and a shortlist. Do this before you look at the click column, or you will label to fit the result.
4. Compare group CTR and group position. If the better-ranking group has the worse CTR, you have found it.
5. Spot-check the SERPs. Pull the live results page for two queries per group and note whether an overview appears and how completely it answers.

If you want the last step done inside your agent rather than by hand, ContextBolt SEO runs it as a serp_overview call and returns the page features alongside the top ten, which is how the table above was built. The AI visibility tools that track citations answer a different question, one worth asking separately.

What should you actually do about it?

Three things, in order of how much they are worth.

Stop pricing informational rankings as traffic. A position-6 ranking on a high-compression query is a brand asset, not a channel. The same caution applies to the metric being sold as the replacement, since prompt volume is modeled rather than counted and cannot be multiplied into a traffic forecast either. It still earns citation, and Seer’s data shows cited brands hold 0.70 percent against 0.52 percent for uncited ones, so it is not worthless. It is just not 500 visits a month, and any forecast that says otherwise is wrong.

Move your writing budget toward low-compression questions. Procedures, comparisons, anything with real trade-offs, anything where the honest answer is “it depends and here is why”. Our Group B is not clever. It is just made of questions that do not fit in a box.

Add something to the high-compression pages that the summary cannot carry. A tool, a table someone will want to sort, a downloadable file, a calculator. Not more words. The summary already beat you on words. It cannot beat you on a thing that does something.

What we are doing on this site is the second and third. The where are bookmarks on X page still gets 66,869 impressions a month and always will, so it now carries a route to the thing Google cannot summarize, which is a search across your own saved posts. That is a worse business than 2019 organic traffic. It is a better one than complaining about it.

Where this data is weak

Four caveats, because a data post that hides its holes is marketing.

Fifteen pages is a small sample. The impression base is large, 118,408, but it comes from one site in one niche. Treat the direction as solid and the exact 79 percent as ours, not yours.

The grouping is a judgment call. We split by answer shape and made the call before looking at clicks, but it is still a human labeling a spectrum as two buckets. Reddit Premium vs free was the hardest one to place.

Four live SERP checks is thin. They are enough to disprove the strict binary version of the theory, which is what we used them for. They are not enough to establish the compression thesis on their own.

Snapshot position and average position disagree. Our page averaged 7.7 for “are twitter bookmarks public” over 28 days but did not appear in the top ten of the live snapshot we pulled on July 31. Both numbers are real. They measure different things, and anyone reconciling a rank tracker with Search Console should expect that gap.

The short version

Rank is no longer a proxy for traffic on informational queries, and the replacement variable is not “does this keyword have an AI Overview”. It is whether the question you answer has a short complete answer.

Our best-ranking pages are our worst-performing pages, by a factor of 4.7, and that is not a ranking problem, a content quality problem, or something a title rewrite fixes. We tried the title rewrite. It moved nothing.

We will re-pull this in October 2026 and publish whatever it says, including if it contradicts this.