Direct answer: Google Search Console gives you no way to filter for AI Overviews, so you cannot measure them directly. You can measure their footprint. On our own property, 2,967 indexed URLs and 875 articles, the share of named-query impressions that earn zero clicks went from 65 percent to 72 percent in three months, and the queries that held their ranking through it were the longest, most conversational ones, not the short commercial ones.
That last part is the finding worth your time. Everyone reports the same headline, that average position is getting worse. On our data, average position getting worse and our commercial pages ranking worse are two different events, and they were happening at different rates to different query shapes. Most dashboards cannot tell them apart, which means most teams are optimizing against a number that is mostly measuring something else.
Why you cannot simply filter for AI Overviews
Search Console documents exactly how AI surfaces are counted. An AI Overview occupies a single position in search results, and all links in the AI Overview are assigned that same position. Impressions follow standard rules, and clicking a link inside the Overview counts as a click. So AI Overview activity is already inside your Web search numbers, blended in, not sitting in a separate bucket.
What does not exist is a searchAppearance value for it. We queried the searchAppearance dimension across four separate 28-day windows on our property. The only value that ever returned rows was TRANSLATED_RESULT. There is no AI Overview appearance type to filter on, in the API or the UI.
Google did add a Generative AI performance report in June 2026, which breaks out impressions for AI Overviews and AI Mode. At launch it carried impressions only, with no clicks, CTR, or query dimension, so it tells you that you were shown and not whether it did anything for you. Useful, additive, and not a substitute for measuring the blend.
What we measured, and on what
One property: a B2B SaaS site with 2,967 URLs in its public sitemap, of which 875 are blog articles. Three windows of exactly 28 days, four whole calendar weeks each, so every window contains four of every weekday and day of week cannot skew a comparison. All pulled from the Search Console API at query grain with type=web and dataState=final:
- Window A, 2026-05-11 to 06-07 (Mon to Sun)
- Window B, 2026-06-22 to 07-19 (Mon to Sun)
- Window C, 2026-07-26 to 08-22 (Sun to Sat, the last 28 days of finalized data)
All figures below are ratios, distributions, and positions. We are not publishing absolute traffic volumes, which would tell you about our size and nothing about the mechanism.
Finding 1: most of what you are measured on is unnamed
Search Console anonymizes rare queries. On our property the named queries, the ones you can actually see and act on, accounted for 25.3 percent of impressions in Window A and 34.8 percent in Window C. The rest is a mass you are graded on and cannot inspect.
You can still characterize it by subtraction, because impression-weighted position is additive. Take the site total, subtract the named rows, and what remains is the anonymized tail:
| Anonymized tail | Window A | Window C |
|---|---|---|
| Share of all impressions | 74.7% | 65.2% |
| Average position | 9.18 | 14.44 |
| CTR | 0.136% | 0.168% |
A CTR near a tenth of a percent across two thirds of your impressions is not a conversion problem. It is a category of impression that was never going to convert, sitting inside the same average you report to your leadership.
Finding 2: the zero-click mass grew, and it moves your headline number
Split the named queries by whether they earned even one click in the window. The two halves behave nothing alike:
| Named queries | Window A | Window B | Window C |
|---|---|---|---|
| Zero-click share of impressions | 65.1% | 74.2% | 71.9% |
| Position, zero-click queries | 9.61 | 12.81 | 14.79 |
| Position, click-earning queries | 9.09 | 10.38 | 8.42 |
Read the bottom two rows against each other. The zero-click mass deteriorated by more than five positions. The click-earning queries ended the period roughly where they started. A single site-wide average blends those two into one number, and because the zero-click mass carries about 72 percent of the impressions, the blended number mostly reports what the zero-click mass did.
The trap this creates. A team watching site-wide average position would have concluded their commercial pages were collapsing. On this property they were not. That is a real decision, made on a real number, pointing the wrong way.
Finding 3: the longest queries held their ranking. The short ones did not
This is the part we did not expect. We froze the query set, taking only the 9,604 queries present in both Window A and Window C, so nothing here is composition change, then split by word count.
| Query length | Queries | Position A | Position C | Change |
|---|---|---|---|---|
| 1 to 2 words | 1,508 | 10.77 | 14.52 | +3.75 |
| 3 to 4 words | 5,836 | 9.97 | 14.54 | +4.58 |
| 5 to 7 words | 1,516 | 8.27 | 11.17 | +2.90 |
| 8 or more words | 744 | 6.25 | 6.53 | +0.27 |
The eight-plus-word queries barely moved, while short and mid-length queries lost three to four and a half positions on the identical page set. And those long queries rank far better in absolute terms, at 6.5 against 14.5.
Look at what they are. In Window A, 10,352 zero-click queries of five or more words carried 26.5 percent of all named impressions at an average position of 7.39. They read like this, verbatim from the report: "who offers the most efficient ai marketing operations automation", "most effective customer data platforms for b2b personalization", "which review automation tools are best for agencies". One of them is a pasted assistant system prompt. These are not humans typing into a search box. They are the sub-queries that AI systems generate when they decompose a question, and they show up in your Search Console as ordinary impressions at strong positions with no clicks attached.
So the composition of our impressions shifted. Machine-shaped queries where we rank well and are never clicked grew as a share, human-shaped commercial queries where clicks live got harder, and the single blended average moved on the arithmetic of that mix as much as on any ranking change.
How this squares with the published research
The direction matches the outside work. Seer Interactive's longitudinal study of 2.43 billion impressions found organic CTR on AI-Overview-triggering queries falling from 1.76 percent to 0.61 percent. Our contribution is the composition underneath that: it is not only that clicks per impression fall, it is that the impression base itself becomes something different, and the queries that survive best in rank are the ones least likely to send a person.
How to measure this on your own property
Five steps, all doable with the Search Console API and no vendor.
- Stop reporting one site-wide average position. Split it: click-earning queries versus zero-click queries. Report the first as your ranking health and treat the second as an exposure metric.
- Pin your aggregation basis and never mix. On the same days, our property reads about 1.7 positions worse under
autoaggregation than underbyPage. Two teams quoting "average position" from the same property can differ by more than a page of results purely on this setting. - Segment by query word count. One extra dimension. It separates AI-shaped fan-out from human demand better than any keyword list, and it needs no classifier.
- Subtract to size the anonymized tail. Site totals minus named rows gives you its impressions, position, and CTR. Two thirds of your grade may live there.
- Never read the last three days. Under
dataState=allthe freshest rows keep growing as they finalize, so a trailing partial day plotted as a finished one bends every trend line down at the right edge. Usefinal, and drop the tail regardless.
One caveat on your own baselines. Google's data anomalies record documents a Web Search logging error affecting impressions, CTR, and average position from 2025-05-13 until 2026-04-27. Any comparison spanning that boundary is comparing two measurement regimes. We tested our own series for a step at the fix date and found none, but test yours before trusting a year-over-year read.
What this changes about the work
If a large share of your impressions comes from machine-generated sub-queries that never click, then ranking for them is not traffic, it is eligibility to be cited. The output metric for that surface is citation presence in AI answers, not sessions. Meanwhile the human commercial queries, the short ones, are where clicks still exist and where our data shows the real position loss happening. Those two need separate targets, separate reporting, and separate work.
The practical failure we keep seeing is a team averaging both into one KPI, watching it fall, and rewriting the wrong pages. Splitting the number costs one afternoon of API work and it changes what you do next.
Frequently asked questions
Can I see AI Overview impressions in Search Console?
Not as a filter. AI Overview clicks and impressions are blended into your Web search numbers, and no searchAppearance value exists for them. The Generative AI performance report added in June 2026 breaks out impressions for AI surfaces, but at launch it carried no clicks, CTR, or query dimension.
Does an AI Overview citation help or hurt my average position?
Mechanically it can improve it. Google assigns every link in an AI Overview the same single position, often a high one, so being cited can pull your reported average toward that position while sending very few clicks. This is one reason average position and clicks can move in opposite directions.
Is a site-wide CTR under 0.2 percent a problem?
Not necessarily, and not on its own. Check the composition first. On our property most impressions sit on queries that were never going to be clicked, which drags the blended rate far below what the click-earning segment actually achieves. Judge CTR inside the segment you are trying to win.
Why would long conversational queries rank better than short ones?
Long queries are more specific, so far fewer pages genuinely answer them and the competitive field is thinner. On our frozen query set the eight-plus-word group sat at position 6.5 while one-to-two-word queries sat at 14.5. The tradeoff is that these queries rarely produce a click.
What do I need to run this analysis?
Search Console API access and a way to hold several 28-day windows side by side at query grain. The analysis is simple; keeping the windows, the aggregation basis, and the data state consistent is the hard part, which is the case for joining any two marketing sources.
