The report answers one question well: which canonical pages received measurable exposure in Google's generative Search features, in which countries, on which devices, on which dates. Treat any claim beyond that as inference, and label it as inference.
What the report contains
Google's documentation is specific about the rollout and the dimensions. Checked directly on 2 September 2026:
As of August 31, 2026, we've rolled out these insights to all websites worldwide.
The documented dimensions are four:
Grouped by final URL, with data assigned to the page's canonical URL rather than a duplicate.
Grouped by the origin location of the search.
Desktop, tablet or mobile.
Grouped by day, week or month, in Pacific Time.
It reports impressions from AI Overviews and AI Mode, top-performing pages, and trends over time, with export.
The dimension that is missing
There is no query dimension. Google does not document one.
That single absence should change how this data gets presented, because the obvious temptation is to bridge it with the ordinary Web performance report, which does have queries. The two are separate observations and Google supplies no join between them.
A query appearing in your Web report does not establish that it triggered an AI Overview or AI Mode impression for your site. It establishes that Google associated a page with that query strongly enough to record an ordinary Search impression. Those are different events.
Relabelling ordinary Search Console query rows as AI prompts is the most likely reporting failure of the next twelve months. It will be common, it will look authoritative in a slide, and it is not supported by the product.
It is also worth knowing what does show up in the ordinary report. Two of our own top ten queries are full sentences, which is visible in the Web view today. If you want prompt-level evidence, you still have to generate it yourself. That means a fixed prompt set with the conditions recorded, which is a different instrument measuring a different thing.
Why canonical attribution matters
Google states it plainly:
Just like the performance report, most performance data in this report is assigned to the page's canonical URL, not to a duplicate URL.
That makes canonical ownership a migration risk, not just a technical hygiene item. If a redesign, consolidation or accidental tag change moves which URL Google treats as canonical, the generative history attaches differently.
The practical consequence is a sequencing one. Export the page-level baseline before anyone touches templates, and record the old and final canonical for every priority URL in the pre-launch handoff register. After launch, that export is the only way to explain what changed.
What to export today
Do this before you need it. There is no back-fill for a baseline you did not take.
- Generative AI report: page, country, device, date, impressions, the date range selected, any filters applied, and the export timestamp.
- Standard Web report, separately: query, page, clicks, impressions, CTR, average position, country, device, date.
- The current canonical for every priority page.
Name the files so the two can never be confused later:
2026-09-02-gsc-generative-ai-pages-28d.csv 2026-09-02-gsc-web-queries-28d.csv
The naming sounds trivial. Six months on, when someone builds a report from whichever CSV was open, it stops being trivial.
Bing measures something different
Bing Webmaster Tools added an AI Performance report in public preview on 10 February 2026. It exposes total citations, average cited pages, page-level citation activity and grounding queries, which are the phrases Bing used when retrieving cited content.
Notably, Bing publishes its own warning about interpretation:
This reflects how often pages are cited, not page importance, ranking, or placement.
So the two platforms give you genuinely different evidence, and neither gives you a ranking:
| Evidence | Google generative report | Bing AI Performance |
|---|---|---|
| Generative impressions | Yes | Not the headline metric |
| Page dimension | Yes, by canonical | Yes, page-level citation activity |
| Country and device | Documented | Not the reported dimensions |
| Query or prompt | Not documented | Sampled grounding queries |
| Citation observation | No | Yes |
| Recommendation or lead | No | No |
Bing gives you sampled query phrases that Google does not. Google gives you impression volume that Bing does not. Use both, and do not average them into one number.
What it does not establish
- Not a ranking. An impression records that content was presented under Google's reporting rules. It is not a numbered position inside an answer.
- Not a recommendation. A page can support a factual passage without the answer recommending the business on it. A brand can be recommended while a different site is cited.
- Not the prompt. Four dimensions are documented and query is not among them.
- Not causation. Impressions rising after an edit is useful timing evidence, not proof the edit caused it. Demand, rollout, crawling and other site changes move too.
- Not a measured zero. Google notes the report may be unavailable if a property lacks sufficient impressions or is excluded from these features. Report unavailable means evidence unavailable.
That last one matters most for small sites. An empty report is not a finding.
A 28-day routine
Day 0. Export both Search Console reports and Bing AI Performance if available. Freeze the pages, canonicals and conversion events being evaluated. Record the publication or launch timestamp.
Within 24 hours. Confirm the live page and its self-canonical, check indexability, confirm internal links and sitemap inclusion, verify the analytics event fires once. This is implementation QA, not a performance verdict.
Day 7. Check crawling, indexation and canonical selection. Fix implementation failures before judging content.
Day 28. Compare page-level direction in standard impressions and clicks, generative impressions by canonical page, Bing citation activity, and qualified conversions. Record the confounders alongside.
Use longer windows when the site has little data. A move from one impression to two is a 100% increase and commercially meaningless. We would rather report a flat 28 days honestly than a percentage that flatters a rounding error.
The sentence to use in a report
This one is defensible:
During this period, Google reported generative AI impressions for these canonical pages. Bing separately reported these cited pages and sampled grounding queries. Neither report by itself shows an AI ranking, a recommendation, or a resulting lead.
This one is not:
"We ranked in AI and generated citations."
The first sentence is less impressive and more accurate. Given that the second cannot be supported by either product, the first is also the only one you can defend when a client asks how you know.
The report makes AI visibility more measurable. It does not make it simple, and it does not make the old habit of turning impressions into rankings any more honest than it was.