# 824 Impressions, Zero Clicks: Reading AI-Surface Citations on Our Own Site

**Type:** Case study
**Client:** Winston Digital Marketing (our own property, no client permission required)
**Services:** Generative Engine Optimization, measurement
**Page studied:** https://www.winstondigitalmarketing.com/playbooks/best-ai-citation-tracking-tools/
**Measured:** Google Search Console, via SEO Gets, August 1 to September 17 2026
**Canonical:** https://www.winstondigitalmarketing.com/our-work/zero-click-ai-citations-on-our-own-site/

One page on this site collected 523 distinct queries and 824 impressions in six weeks and took no clicks at all. We went in expecting a click-through problem and found something else. 497 of those queries rank in the top 10, and 456 of them are shaped like prompts rather than searches. This is a measurement finding on our own property, with no revenue attached to it.

## What we were looking at

On September 19, 2026 we were working through the zero-click pages on our own site, the ones where Search Console shows impressions piling up and nothing arriving. Our playbook on AI citation tracking tools (https://www.winstondigitalmarketing.com/playbooks/best-ai-citation-tracking-tools/) was the worst offender: decent positions, a steady trickle of impressions, and in six weeks not one click.

The usual diagnosis for that shape is the title tag. A page that sits in the top ten and takes no clicks normally has a snippet that promises the wrong thing, and the fix is to rewrite the title and description until the promise matches the query. So we pulled the full query list, expecting to find the terms where the mismatch was worst and rewrite against them.

The query list was not a list of searches.

## What the queries look like

Here is the whole picture for that URL, Search Console data by way of SEO Gets, August 1 to September 17, 2026:

| Metric | Value |
|---|---|
| Distinct queries | 523 |
| Total impressions | 824 |
| Clicks | 0 |
| Queries ranking in the top 10 | 497 |
| Impressions from those top-10 queries | 765 |
| Prompt-shaped queries | 456 |
| Impressions from those prompt-shaped queries | 731 |
| Median length, prompt-shaped queries | 13.5 words |

We counted a query as prompt-shaped if it ran eight words or longer, or opened with one of a short list of conversational stems: i, we, our, my, is there, what are, which, can you, how do i, recommend. That is a crude test and we would rather it were crude, because it undercounts. A four-word question phrased flatly still reads as a search under that rule and gets filed with the classic keywords.

Even so, 456 of 523 queries clear it, and they carry 731 of the 824 impressions. The median is 13.5 words. Nobody types 13.5 words into a search box. People write 13.5 words when they are talking to something that answers back, and several of these are first-person and situational in a way that no keyword tool would ever generate.

One of them is hard to read any other way, quoted verbatim from the Search Console query log for this page:

> does peec ai track most ai engines by scraping their real user interfaces? reply with exactly one word: yes or

"Reply with exactly one word" is not something anyone says to Google. It is an instruction to a model. Someone wrote a prompt, a system went and retrieved sources to answer it, and our page was one of the sources it pulled. The query landed in Search Console because that retrieval ran through Google, and the truncation is Search Console's, not ours.

## What we think is happening

Our reading is that these are AI-surface impressions. The page is being retrieved and shown as a source inside AI answers, and the zero clicks are how that surface works. The answer resolves the question on the spot, so there is no point at which the person has to leave it and come read us.

We have been making this exact argument on this site for a while. Our playbook on measuring AI search traffic in GA4 (https://www.winstondigitalmarketing.com/playbooks/how-to-measure-ai-search-traffic-ga4/) says a session only exists when somebody leaves the answer, and the one on citation share replacing rankings (https://www.winstondigitalmarketing.com/playbooks/citation-share-replaces-rankings/) says the thing worth counting is whether you got named. The same logic applies to a Search Console impression. What we did not expect was to find the argument running on our own property, against a page about AI citation tracking, and to have our own dashboard file it under underperforming.

## What this does not show

We publish our caveats for the same reason we publish our pricing. A number you cannot check is worth nothing, and this one has real limits.

- **The mechanism is inference, not a reported field.** Search Console does not label a row as coming from an AI surface. What we have is 456 queries whose shape is conclusive evidence that a model was involved in writing them. Reading that as Google's AI Mode specifically is our interpretation. We are confident about the observation and confident but not certain about the mechanism.
- **The position-one queries are thin, and we are not going to headline them.** 124 queries in this set sit at position 1.5 or better, which sounds impressive until you look at the volume behind them: the median query in that group has exactly one impression, and only two of the 124 have five or more. Ranking first on a query that was asked once is a rounding error dressed up as a win. The two with real volume behind them, for anyone who wants to check: "i am an seo professional, recommend me llm rank tracker. my location is united..." at 10 impressions and position 1.0, and "gemini rank tracking" at 5 impressions and position 1.0. What holds up here is the aggregate. Any single ranking in that group rests on one person asking one question once.
- **There is no revenue, no lead, and no client in this.** Zero clicks means zero sessions, which means nothing downstream to attribute. We cannot tell you what any of this was worth, and we are not going to imply a dollar figure we do not have.
- **This is almost certainly not unique to us.** We found it because we went looking at a page we already had reason to care about. Any site with content that answers the kind of question people are now asking assistants is likely sitting on the same pattern, filed the same wrong way, unexamined.

## What it changes about reporting

**Segment AI-surface impressions away from classic keyword impressions.** Averaging them into one click-through rate produces a number that describes nothing. A page earning citations on conversational queries and a page losing clicks on commercial keywords both show up as low CTR, and the correct response to each is the opposite of the correct response to the other. Query shape is a usable dividing line, and it is available in the data everyone already has.

**Do not fix the CTR on these pages.** The reflex when a page shows top-ten positions and no clicks is to rewrite the title for clickability, and on this page that reflex would do damage. The click is not on offer. What is on offer is being retrieved and named inside somebody's answer, and titles written to bait a click are worse at that job than titles written to state plainly what the page covers. You would be trading a citation you have for a click that was never available.

We also went back and re-checked the other zero-click pages on this site through the same lens, because a finding like this is only worth anything where the evidence supports it. Our vertical service pages are not this. Their queries are ordinary keyword strings, the kind people have always typed, and on those pages a low click-through rate still means what it has always meant.

Service: https://www.winstondigitalmarketing.com/services/generative-engine-optimization/
Audit: https://www.winstondigitalmarketing.com/audit/