Search Console Is Undercounting. Snapshot the Range Now

Search Console’s Generative AI performance report is under-counting impressions, and Google has confirmed it. A logging error affects data starting August 13, 2026 — still affected as of the August 17 report — and a related error hit the Discover performance report and its Generative AI sub-slice for August 13 data. This is a data defect, not a visibility change in Search.

The timing could hardly be worse for clear thinking. Google’s August 2026 spam update began rolling out on August 18 — a genuine ranking event, landing five days after a reporting defect began suppressing impression counts. Anyone glancing at a dashboard this week risks blending the two into one story, and the two stories demand opposite responses.

This playbook covers exactly what Google has confirmed, how to keep the logging bug and the spam update strictly separate, what an earlier 50-week logging bug teaches about how these entries get closed out, and the concrete mechanics — UI and API — for snapshotting the affected range before the picture changes again.

Google’s own data-anomalies log — the official registry of Search Console data defects — carries two related entries. The first: a logging error caused a decrease in impressions on the Generative AI performance report in Search, for data starting August 13, 2026, still affected as of the August 17 report. The second: a logging error decreased clicks and impressions on the Discover performance report for August 13 data — Discover’s own Generative AI sub-slice included. If you track AI surfaces through the Generative AI performance report, the dip you see from August 13 is the bug, not your content.

Both entries carry the same clause in Google’s institutional voice: “This issue affects data logging only.” That is Google’s standard phrasing for “the numbers are wrong, the visibility is not” — an independent confirmation that lines up with what John Mueller said publicly, as reported by Search Engine Land on August 17.

Two framing cautions before anything else. First, the anomaly entry is a live document: Google updates it as the issue evolves, so don’t treat any currently displayed end date as final — the reliable fact is that data starting August 13 is affected. Second, this isn’t the first data-quality quirk on the AI reporting surfaces: Search Console’s AI Mode data has already had its share of quality quirks, which is part of why a confirmed anomaly entry — rather than community speculation — is the signal worth acting on.

On August 18 — the day this post publishes — Google began rolling out its August 2026 spam update, confirmed on its Search Status Dashboard. That is a genuine ranking event: in Google’s own words, “a normal spam update” that “will roll out for all languages and locations,” enforcing the existing spam policies rather than any new rule. Google has not disclosed what share of queries is affected, and its standing guidance is that recovery from spam updates can take many months.

The logging bug is none of those things. It changed no rankings, touched no policies, and began five days earlier — August 13 — on two specific reporting surfaces. The table below is the side-by-side most coverage doesn’t give you, because the two events are usually written up as separate stories.

The practical test is simple: an impressions dip that starts precisely on August 13, confined to the Generative AI (Search) and Discover reports, while clicks on the core Search report behave normally, is the logging bug. Ranking movement that begins on or after August 18 and builds gradually across days is spam-update territory. A step-change on a reporting surface with a dated anomaly entry is an accounting problem, not an SEO problem.

This is not Search Console’s first logging defect this cycle. An earlier logging error inflated impression counts for roughly fifty weeks — from May 13, 2025 until April 27, 2026 — before Google fixed it, as Search Engine Roundtable documented in May 2026. That bug affected impressions, CTR, and average position — not clicks — and ran in the opposite direction from today’s: it over-reported, where the August 13 bug under-reports.

The load-bearing detail is what happened after the fix. Per Search Engine Roundtable’s reporting of Mueller’s Bluesky confirmation, the correction was forward-only: the ~50 weeks of inflated historical data were never restored to accurate values. Google marked the entry resolved, the pipeline logged correctly from that day forward, and the historical distortion simply stayed in everyone’s reports.

To be explicit about what this does and doesn’t establish: Google has said nothing about backfilling — or not backfilling — the current August 13 bug. The no-backfill expectation is an inference from precedent, and should be treated as exactly that. But it is a well-sourced precedent, on the record, about the most comparable recent incident. Planning around it costs you one export; ignoring it risks a permanent hole in your AI-surface baseline.

Here is the operational nuance most coverage of this bug skips. On the one precedent we have, Google marked the anomaly-page entry “resolved” once the logging pipeline was fixed going forward — not once the history was repaired. That entry reads “This issue has been resolved,” for an incident whose historical data, per Mueller’s separate confirmation, was never corrected. Resolution referred to the fix, not to restoration.

That means the moment to act is now, while the affected range is current and clearly bounded on one side — not after a resolution notice appears. When the current entries eventually flip to resolved, the most likely reading, based on precedent, is that new data logs correctly from that day forward while the August-13-onward dip stays in the report exactly as it looks today.

Even without a bug, Search Console data is not forever. Google’s Performance report documentation states the default view shows the past three months; the full history is available by widening the date-range picker. Beyond that, Search Console’s widely documented ~16-month rolling retention window — industry operating knowledge rather than a stated Google guarantee — means older performance data ages out of the tool entirely. The bug adds urgency to a discipline you arguably needed anyway: keep your own copies.

One honest caveat on the API route: Google’s own documentation for the Search Analytics query endpoint warns that it “does not guarantee to return all data rows but rather top ones” — at very high row counts an export is bounded, not a certified-complete dump. That doesn’t weaken the advice; it sharpens it. A top-rows snapshot taken this week is still categorically better evidence than a report whose affected range may never be corrected.

What should the snapshot cover? At minimum: the Generative AI (Search) report and the Discover report from August 1 through the most recent available date, so you hold clean pre-bug days ahead of the August 13 boundary. If you can afford the rows, pull the core Search performance data for the same window — it gives you the unaffected control series that makes the bug’s shape provable later.

The scenarios below cover the four dashboard patterns most teams will meet this week. The connecting thread is one instinct: knowing when a Search Console signal needs action versus a shrug — and this week, a confirmed anomaly entry is the difference between the two.

Step back and the trend is worth naming: this is another confirmed Search Console logging defect within months of the last major one, and this one hits the Generative AI reporting surfaces specifically. Those surfaces are young — the pipes carrying AI Mode and AI Overviews data into Search Console are among the newest in the tool — so it shouldn’t surprise anyone that they’re where the logging faults surface. Measurement infrastructure for AI search is still being built underneath us, in public.

Looking forward, the reasonable projection is more of these, not fewer, as Google keeps extending AI-surface reporting — which means the durable answer isn’t reacting to each anomaly entry but building an export cadence that makes any single gap survivable. That’s the measurement layer we build in our analytics engagements, and it’s becoming inseparable from agentic SEO work generally: you cannot optimize AI-surface visibility on data you don’t reliably hold.