One Page, 64,191 Impressions, Zero Clicks in 87 Days

One Page Took 64,191 Impressions and Zero Clicks in 87 Days

Between 2026-06-13 and 2026-09-07, using only complete reporting days, Google Search Console recorded 64,191 impressions and zero clicks for one page on this site, /topics/anthropic-agent-sdk. Five of those 87 days recorded no impressions at all. None of them recorded a click.

The cause is unresolved. A large impression count with no clicks does not by itself establish AI activity, a reporting artefact, or a page that is quietly succeeding. This article reports what the data shows, marks where the evidence stops, and describes the checks that would narrow it.

It also corrects a mistake we made along the way, because that correction turned out to be the most useful finding here.

A note on evidence and timing. Everything below comes from exports saved alongside this article. Where a figure uses Google's preliminary same-day data, we say so. The 64,191 total uses complete days only; the equivalent figure including two preliminary days is 68,439, and the two reconcile exactly.

What Does Google Actually Count as an Impression?

An impression records a link a user saw or could have seen. Google's Search Console counting rules, checked 2026-09-10, generally count a link on the current page of results even if the user never scrolls to it. Inside carousels and expanding widgets, the item typically has to enter view or be expanded before it counts. For AI surfaces, the same documentation states that standard impression rules apply to both AI Overviews and AI Mode.

Those rules do not document background retrieval as a counted impression. An explanation resting on invisible machine activity therefore needs evidence beyond a large impression count, and we do not have it.

Why a position near 2 does not mean the second blue link

Google assigns an AI Overview a single position and gives every link inside that block the same position. A query row near position 2 could describe such a link rather than the second conventional result.

That is a limit on what position tells you, not evidence that our page appeared in an AI Overview. Query-level positions also describe different observations from a page average. Our page averaged 4.0 across the most recent window while individual query rows sat near 2, and only the page average describes the page.

What Is Query Fan-Out, and What Does It Not Explain?

Query fan-out is a technique Google may use to assemble an AI response. Its AI features documentation, checked 2026-09-10, describes issuing related searches across subtopics and data sources, and states that AI-feature activity is included in Search Console under the Web search type.

Fan-out describes how the system gathers candidates. The impression rules describe how links shown in results are counted. Neither document establishes that a background sub-query creates an impression or surfaces as its own query row.

Our assessment, labelled as ours: the rows below do not identify their origin. Fan-out is one possible explanation worth investigating. We have no evidence that ranks it above the alternatives, so we are not going to write as though we do.

What the Anomaly Looks Like

The saved export holds 3,153 unique query rows for 2026-08-14 to 2026-09-10, carrying 91,925 impressions and 93 clicks. Those are the rows returned by our paginated requests, not a complete account of searches involving the property.

Set against the sitewide totals, that export covers 67.02 percent of impressions and 18.75 percent of clicks. Google withholds anonymised queries and separately warns that requests grouped by query can drop data. Pagination does not remove those limits, so every share below describes the extracted rows only.

Within the extract, 3,103 of 3,153 rows produced no clicks, accounting for 96.2 percent of extracted impressions.

Two kinds of long query, both earning nothing

The long rows are not one population. Some are keyword-style permutations that stack the same technical terms in different orders. Others are conversational full-sentence questions. All five rows below are property-level counts read straight from our Search Console query export for 2026-08-14 to 2026-09-10. That export is held internally rather than published, so these figures are reported rather than independently checkable. The extraction settings behind them are described in the audit steps further down.

Both styles earn effectively nothing. Neither style records who issued it. People type keyword lists, and automated systems generate natural-sounding questions, so wording cannot separate the two. That is why this article gives you no word-count rule for classifying impressions. Our own export refuses to support one.

The Decline Came From Somewhere Else Entirely

The zero-click page explains none of our lost clicks. It recorded zero clicks in both comparison windows, so arithmetically it could not have lost any, and the decline sits entirely on other pages.

That is the correction, and it is the part worth your time. Our first instinct was to use the zero-click page to explain why our clicks were falling. Across matched complete windows, 2026-08-11 to 2026-09-07 against 2026-07-14 to 2026-08-10, clicks fell from 965 to 576, a loss of 389, while impressions rose 10.1 percent.

The zero-click page recorded zero clicks in both windows. It contributed none of the loss. Its impressions grew from 3,584 to 47,452 over the same period, which is why it dominates any impression chart and why it was so tempting as an explanation.

Here is where the clicks actually went.

Four pages account for 257 of the 389 net lost clicks, which is 66.1 percent. Three of them are free tools. The cause of that decline remains unresolved.

That last sentence is doing real work. The diff establishes location, not cause. AI-related changes could perfectly well be affecting those tool pages, and so could a ranking change, a seasonal dip, a competitor, or a change in how results are presented. Click counts cannot separate those, and we have not yet run the checks that would.

The method matters more than our particular result, because it is the step we skipped. Pull clicks by page for two equal windows that both end on a complete reporting day, join them on the page URL, and sort by the difference. It takes about five minutes. What it gives you is a ranked list of where the decline sits, before any theory has a chance to attach itself to the most eye-catching row in the report.

We did it the other way around. We found the strangest page first, built an explanation that fit it, and only ran the page diff after an external reviewer pointed out that a page with zero clicks in both periods cannot lose clicks. The explanation was elegant and it was about the wrong page.

Four Checks We Ran, and What Each One Could Not Settle

Each check below locates something worth examining. None identifies whether an impression was served to a person or produced by a machine.

Check one: query length. Long rows correlate with zero clicks in our extract. Correlation is not identification, and as shown above the long rows contain both keyword-style and conversational wording with no authorship recorded for either. Use length to segment. Do not use it to conclude.

Check two: reported position against clicks. A row near position 2 with no clicks is unusual enough to examine, but every link inside an AI Overview carries that block position, so the number does not identify the result type. The page average of 4.0 is the figure that describes the page.

Check three: the daily shape. The page produced 1,745 impressions on 2026-06-19, then 15 on 2026-06-26 and single digits for a week. It went to zero for five days in late July, then climbed from 297 on 2026-08-17 to 3,924 on 2026-08-31 and held above 3,000 through 2026-09-07. Bursts like that are consistent with automated tooling, with a staged feature rollout, and with other explanations we cannot separate.

Check four: device and geography. This produced the most surprising number in the investigation, and we had not run it until an external review pushed us to look past the query data.

Every device-classified impression was desktop. The API returned no mobile rows and no tablet rows at all for this page and window.

Read those shares carefully. Both the device and country extracts total 63,777, which is 414 short of the 64,191 in the daily series, a gap consistent with rows being dropped when results are grouped. So 414 impressions carry no device or country classification here, and the percentages above describe the segment extract rather than the headline total. Five countries cover 99.6 percent of that extract.

Even scoped that tightly it is an odd distribution. A general-audience page normally shows a substantial mobile share and a longer geographic tail. We are still not going to tell you it proves machine origin, because a narrow desktop-only technical audience is entirely possible for a page about a software development kit. What the check does is narrow the field, and it is the strongest single reason we are still investigating rather than closing the file.

What We Ruled Out, and What We Could Not

No report Google offered as of 2026-09-10, the date we checked its documentation, resolves this.

Search Console does have a Generative AI performance report covering AI Overviews and AI Mode, grouped by page, country, date and device. It has no query dimension. The Search Analytics API has no AI Mode or AI Overviews search type at all.

The strongest available check is a page-level comparison: filter that report to the page and compare its AI-feature impressions against the same page in the standard Performance report, matching dates and aggregation. That can establish reported AI-feature exposure. It cannot identify fan-out queries and it cannot explain a change in clicks. We have not completed it, and we will not describe its result before we have.

Explanations we cannot exclude:

1. Automated tooling. Rank trackers and scrapers generate impressions and would produce a similar burst pattern.
2. AI-feature exposure. Consistent with the position behaviour, but unconfirmed until the page-level comparison runs.
3. A real audience that does not click. Harder to reconcile with 87 days and no clicks, but not excluded.
4. Reporting artefacts. Impression counting for AI surfaces is new and still rolling out.

We are not ranking these. We do not have the evidence to.

What Zero Clicks Does and Does Not Tell You

Zero clicks alone identifies neither failure nor success. Interpreting it requires knowing what was searched, what result appeared, who was searching, and what the page was for.

Pew Research Center, publishing 2025-07-22 from browsing data contributed by 900 US adults in March 2025, found people clicked a traditional result on 8 percent of visits where an AI summary appeared against 15 percent where none did, and clicked a source inside the summary on 1 percent of visits.

That establishes that clicking behaviour differs between those two situations. It does not establish the cause of any particular site's decline, including ours. Treat it as context for why a click-through rate can fall without a page changing, not as a diagnosis you can apply to your own dashboard.

Two related findings are worth holding alongside it. Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande defined generative engine optimization in November 2023 around inclusion in a synthesized answer rather than selection from a list. And Venkit, Laban, Zhou, Mao and Wu evaluated answer engines across 303 queries in 2024, finding substantial gaps in citation support. Inclusion and accuracy are separate outcomes, and only the second one protects a brand.

What We Are Doing Next

Running the page-level AI comparison. It is the only check that would establish AI-feature exposure, and it comes before any further theorising.

Investigating the tools decline as its own problem. Four pages lost 257 clicks. That deserves its own evidence rather than being folded into an AI narrative because an AI narrative was already on the page.

Separating sampling frequency from reporting cadence. Schulte, Bleeker and Kaufmann found source sets overlapping by only 34 to 42 percent between consecutive days, and recommend at least 7 runs per prompt per day with rolling aggregation over two to four weeks. Those are their recommendations, not a universal law, and any report we publish should disclose its prompt set, observation count and measurement dates.

What This Means If You Are Seeing the Same Pattern

Rising impressions with falling clicks is common enough right now that it is worth saying what this investigation does and does not license you to conclude.

It does not license a diagnosis. If your report looks like ours, you have the same evidence we have, which is not enough to name a cause. The honest statement to a stakeholder is that impressions rose, clicks fell, the two are not necessarily connected, and you are running the page diff to find out where the loss sits. That is a smaller claim than most dashboards imply and it is the one you can defend.

It does license two concrete actions. Run the page-level click diff across matched complete windows, because it costs minutes and tells you which pages the decline sits on. Then run the Generative AI report comparison for any page with a large impression count, because it is the only check available that speaks to AI-feature exposure at all.

It should change what you promise. A team that reports "AI is taking our clicks" without a page diff is guessing, and if the guess is wrong, whatever is actually happening keeps happening. In our case the decline was sitting on four tool pages nobody was watching, precisely because a more interesting page was absorbing the attention. We still do not know why those pages fell. That failure mode does not require AI search to exist. AI search just gave it a more persuasive story to hide behind.

Watch your own segments before you accept a narrative. The device split was the check that most changed our thinking, and we nearly did not run it. If a page shows an implausible segment distribution, whether that is 100 percent of one device, one country, or one browser, you have learned something about the traffic that no amount of query analysis would have told you. Segment checks are cheap and most audits skip them.

How Do You Run This Investigation Yourself?

1. Pull complete days. Set data_state to final. Same-day data is preliminary, and comparing a preliminary window against a historical one manufactures a difference.
2. Use matched windows. Equal length, both ending on a complete day. Ours were 28 days each.
3. Export every row. The API accepts a rowLimit up to 25,000 per request with startRow pagination. The default of 1,000 is where partial exports get mistaken for complete ones.
4. Record your coverage. Compare your export totals against the property totals and publish both figures. Ours covered 67.02 percent of impressions and 18.75 percent of clicks.
5. Diff clicks by page across the two windows. This is the step that tells you where a decline actually came from. Run it before you form a theory, not after.
6. Segment by query length to locate anomalies. Then stop, because the segmentation does not identify origin.
7. Compare the page in the Generative AI report. The only check that speaks to AI features, and only at page level.
8. Compute percentages from unrounded totals. Rounded rates produce deltas that will not reconcile.

The Part Worth Keeping

One page, 64,191 impressions, 87 days, zero clicks, cause unresolved. We can tell you precisely what happened and precisely how far the evidence reaches, which is less than we wanted when we started.

The more useful lesson is the one that cost us a draft. The most dramatic number in a report attracts every explanation, and it was the wrong page. Our lost clicks came from four ordinary tool and guide pages while our attention was on a topic page that recorded no clicks during the observed period. Diff your pages first. The interesting anomaly and the expensive problem are often not the same thing.

Conclusion

We found a page with 64,191 impressions and no clicks, could not explain it, and published the investigation rather than a theory. The measurement surface for AI search is genuinely incomplete: no Google report connects a query row to an AI feature, so anyone offering you a clean split between human and machine impressions is inferring from the same partial data you have.

The finding we did not expect is the one we would act on first. The page that looked alarming recorded no clicks during the observed period, and the pages that quietly lost hundreds of clicks were ordinary tools we were not watching. Why they fell is still open. Diff your pages across matched complete windows before you build a narrative, because the narrative will fit whatever page you were already looking at.

We traced earlier parts of this shift in Google Traffic Dropped 33% and Our Biggest AI Referrer Cites Us the Least. Both are worth reading against this one, including where they now look more confident than the evidence supported.

Want to check the same things on your own site?

• Radar - Audit how AI engines read your brand in one pass, with a fix attached to every check.
• AI Citation Tracker - See whether answer engines cite you, and what they say when they do.
• AI Crawl Checker - Confirm the crawlers behind AI features can reach your content at all.
• Contact Us - Send us your Search Console export and we will run the page diff with you.