Which of Your Impressions Are AI Agents? A 20-Minute Test
Google folds AI Mode into your Performance report with no filter. A 20-minute test that separates prompt-shaped agent rows from people who could click.
A meaningful share of the query rows in your Search Console Performance report were never typed by a person looking at a results page, and Google gives you no filter to find them. You can separate them yourself in about twenty minutes: pull 28 days of rows, flag the ones carrying prompt shapes no human types, then check whether clicks arrived where the recorded position says they should have. Rows that are prompt-shaped with zero clicks from the top of page one are agent impressions. They are not a snippet problem, and rewriting your title to fix them is wasted work.
What Google now reports, and what it will not separate
Google's documentation on AI features is unusually direct about the reporting boundary. Sites appearing in AI Overviews and AI Mode are counted in the overall search traffic in Search Console, reported inside the Performance report under the "Web" search type. When Google added AI Mode to those totals, the changelog sentence was one line: data from AI Mode now counts toward the totals. Search Engine Journal covered the change and flagged the part that matters if you read CTR for a living — it is mixed into existing numbers, with no dedicated tab and no surface dimension.
Two details in that documentation do most of the diagnostic work:
- Follow-up questions start new queries. Google states that when someone asks a follow-up in AI Mode, all impression, position and click data in the new response is counted as coming from that new user query. A five-turn conversation becomes five query rows.
- An impression only requires that the link could have been seen. Google's definition of impressions and position says an impression counts when someone saw a link to your site, and that depending on the result type, the link might need to be scrolled or expanded into view. The same page makes an argument most reporting overlooks: aim for meaningful impressions, not more of them.
Put those together and the arithmetic gets uncomfortable. A page included as a supporting source can register an impression while sitting in a citation block nobody expanded. That is a real impression by Google's definition and a zero-click row by yours. Multiply it across a model that fans one question out into a dozen parallel searches, and your CTR denominator becomes a mix of two populations — people who could have clicked, and retrieval passes that never had a click to give.
Three row shapes that no human types
You do not need a vendor tool. You need to read your own query list and sort it. Almost every contaminated row falls into one of four shapes.
| Shape | What the row looks like | Where it comes from |
|---|---|---|
| Sentence-length prompt | Lowercase, full question, punctuation, a brand and a constraint in the same string — "what does a good cpm look like for a b2b saas account in mexico" | An AI Mode prompt, or a synthetic query fired by a tracking tool |
| Embedded operators | -filetype:pdf, -site:reddit.com, OR chains, or a string carrying a dozen filters at once | A prompt set or research script. No one types twenty filters into a search box |
| Follow-up fragment | "and the second one?", "where did you find this", "is that in europe too", "how much is it" | A conversation turn in AI Mode, counted as its own query |
| Identical repetition | The same long string, same impressions, several days running | A scheduled run. A human asks roughly once |
That first row is the one people misread as demand. A forty-word question with a geo modifier and a brand name looks like a dream query in a keyword report. It is a prompt.
The residual test
Here is the whole procedure. It takes one export and a spreadsheet column.
- Export 28 days, Queries dimension, Web search type, with Clicks, Impressions and Position. If your volume is high, filter to one page at a time — the row list is capped, and a page-level pull gets you the rows that matter for that page.
- Tag every row with the shapes above: agent-shaped, human-shaped, or unsure. For a week's work, take the top 100 rows by impressions. That is where the distortion lives.
- Set a deliberately low CTR floor for the positions involved. Even 5% for positions two through four is conservative against any published curve. Do not borrow a benchmark curve as truth here — a curve built from Search Console data inherits the same contamination you are trying to measure, so treat it as an upper bound. Advanced Web Ranking's monthly curve is honest about being built from Search Console; that is exactly why it can only be a ceiling in this exercise.
- Compute expected clicks. Impressions multiplied by your floor. Then compare to actual clicks on the same rows.
- Read the gap by group, not in total. Agent-shaped rows with zero clicks at positions one to four and human-shaped rows converting at or above your floor in the same window is the signature. Missing clicks from both groups is a different disease.
A worked example, with numbers I am inventing so the arithmetic is visible: 4,000 impressions sit at an average position of 3, with zero clicks. At a 5% floor, that position owed you 200 clicks. Two explanations fit. Either your snippet lost 200 people, or 4,000 people were never standing there. Open the query list. If those impressions concentrate in prompt-shaped rows, the second explanation wins, and no title rewrite will move a single one of them.
Three diagnoses that look alike at the page level
This is where most audits go wrong: a low CTR at a good position gets one label, "snippet problem," and the fix is a rewrite. There are at least three distinct failures behind that number, and they need opposite responses.
| Signal | Agent noise | Snippet problem | Wrong-intent traffic |
|---|---|---|---|
| Query shape | Prompts, operators, fragments | Short, common, human | Short, human, different job |
| Clicks at positions 1–4 | Zero across many rows | Present but under the curve | Present, weak engagement |
| Page-level CTR | Low and steady | Low, moves when you test a title | Middling CTR, high bounce |
| What actually fixes it | Nothing. Exclude it from the denominator | Title, meta, page format | Intent match or a new page |
The middle column is the one you have been treating as the default. The left column is the one that has quietly been growing for two years, and it is the only one where a rewrite is provably useless.
What to do with each class of row
Ignore agent rows in reporting, and say so. Remove them from your CTR denominator before you judge any title change, and write the exclusion into the report. "Impressions, excluding prompt-shaped rows" is one line, and it changes decisions more than another month of A/B testing titles against a contaminated baseline.
Do not chase the prompt. A citation in an answer with no click available is not a click you lost. If you want that surface, the work is being citable and quotable — a different job, closer to what we described in why crawlers can fetch you without citing you.
Do fix human-shaped rows at positions four through ten with a CTR under one percent. Those are real people who read your title and chose something else, and the standard fixes apply.
Read operator strings as intelligence. A row carrying fifteen -site: filters means somebody — an agency, a competitor, a research desk — is running a prompt set against your category. It tells you which comparison questions the market is asking. It does not tell you what to publish.
What this test does not prove
Three limits, stated plainly, because the temptation here is to turn a diagnostic into a statistic.
- You cannot attribute a row to an engine. Google does not label the surface. AI Overviews, AI Mode and a third-party tool hitting the live results page can produce rows that look identical in the export.
- The row list is a subset, not a population. Low-volume and privacy-protected queries are dropped or bucketed, so your classified rows will not reconcile to your totals. Report the classification as counted versus uncounted, not as a precise percentage of all traffic.
- It says nothing about revenue. There is no session to join, because the click never happened. Anyone who attaches a conversion number to agent impressions is guessing.
That last point is not a reason to skip the test. It is the reason to keep the conclusion narrow: you are fixing a denominator, not discovering a channel.
The report change worth making this month
If your CTR has been sliding at a stable position and every snippet rewrite has failed to move it, stop rewriting before you run this test. In our client work the pattern repeats: a page holds strong positions, the query list is full of prompts and follow-up fragments, and the team has already spent a quarter tuning titles for an audience that was never there.
The fix is unglamorous. Split the denominator, report both numbers, and put your editing hours on the rows where a human was actually reading. If you want the version that also covers prompt-set measurement and citation tracking, our GEO and AI search service is where we run it, and the prompt-set framework shows how to build the query set the second half needs.
Frequently Asked Questions
Can I filter AI Mode traffic in Search Console?
No. Google includes AI Mode clicks and impressions in the Performance report totals, inside the Web search type, and there is no surface filter, no separate row and no dimension that isolates them. Any split has to be built by classifying the query rows yourself.
Why would an AI answer create an impression with no click?
Because an impression counts when a link to your site was seen or could have been seen. In a generated answer the supporting link can sit below the fold or inside a collapsed citation block. The impression is real, and the click was never available.
Does a low CTR always mean my snippets are weak?
No. Before rewriting titles, check what produced the impressions. If a cluster of prompt-shaped queries holds zero clicks at positions one to four, the denominator is contaminated and the snippet was never the problem.



