For two years the honest answer to "how would we know if AEO worked?" was: partially, and with caveats. That is still the honest answer. It just got a little better on one surface.
What Google actually shipped#
On June 3, 2026, Google announced Search Generative AI performance reports in Search Console, with dedicated views for Search and Discover.
Per the announcement, the reports show a separate cut of visibility inside generative AI features such as AI Overviews and AI Mode:
- Impressions — how often your URLs appeared in those features
- Pages — which specific URLs appeared
- Countries — visibility by market
- Devices — for Search results
- Dates — hourly, daily, weekly and monthly granularity
The data also remains inside the overall performance report; this is a dedicated view of it. Google noted the reports rolled out to a subset of websites first, and that additional metrics are under consideration.
That is a genuinely useful addition. It is also narrower than the headlines suggested.
What it does not cover#
Four limits worth stating plainly before anyone builds a dashboard on it:
1. It is Google only. ChatGPT, Gemini, Perplexity, Copilot and the assistants embedded in phones, browsers and business software each retrieve differently. None of them send you a report. For most B2B and considered-purchase categories, a meaningful share of AI-assisted research now happens outside Google entirely.
2. It is impressions, not inclusion quality. An impression tells you a URL appeared. It does not tell you whether you were the recommendation, a supporting link, or the counterexample.
3. It is not clicks, and clicks are not revenue. Pew measured link clicks inside AI summaries at 1% of visits in its July 2025 study. Reading this report as a traffic forecast will disappoint you.
4. Availability varies. Rollout is partial, and the definitions may change as Google iterates.
"New data is easy to over-read. Use it, label where it came from, and keep testing the engines that report nothing."
Andrew Bethel — COO, Perfectus Labs
How to measure the rest#
No single source covers AI visibility today, so treat measurement as a portfolio rather than a dashboard:
- First-party surface data where it exists — the Search Console report, exported and baselined.
- Prompt-level testing across the engines your buyers actually use, run on a fixed schedule with a fixed question set so results are comparable over time.
- Referral and assistant traffic segmented in analytics, so AI-sourced sessions are not buried in direct.
- Demand-side signals — branded search volume, direct traffic, and the "how did you hear about us" answer on your forms. When a model names you and the user never clicks, this is where it shows up.
None of these is authoritative alone. Together they are enough to tell whether the work is moving anything.
Share of answer, not share of clicks#
The useful frame is share of answer: how often you are named at the moment of decision, across the engines your market uses. Impressions inside Google's AI features are one input to that. So is a Perplexity answer that recommends you by name to a buyer who never visits your site.
Measure it alongside pipeline, not instead of it.
How we handle this in Cognos#
Cognos reconciles first-party surface data with scheduled prompt testing across multiple engines, then ties movement back to the levers that caused it — crawl access, entity consistency, payload structure, and content only you can publish.
We report Google's data as Google's data, and everything else with its sampling method attached. A client should always be able to see where a number came from.
What to do now#
- Check whether Search Console has enabled the reports for your properties, and export a baseline. You cannot recover history you did not save.
- Build a fixed prompt set for the non-Google engines your buyers use, and run it on a schedule.
- Segment assistant and AI referral traffic in analytics before it disappears into direct.
- Ask any AEO vendor which of their claims can be reconciled against a source you control.
The measurement excuse is weaker than it was. It is not gone.
Sources: Google Search Central, Jun 3 2026; Pew Research Center, Jul 22 2025. Report availability varies by property.
