Search Console Now Shows Where AI Cites You, and Almost Nothing Else.

Google added Generative AI performance data to Search Console, a view of how often your URLs appear in AI Overviews and AI Mode, alongside generative features in Discover (Google Search Central). For the first time, you can see which of your pages are surfacing inside Google's AI answers. It is also one of the thinnest reports the tool has ever shipped.

The report carries five dimensions. You get impressions, pages, countries, devices, and dates, down to hourly granularity. It then withholds the four numbers an SEO reaches for first. There is no click data, no click-through rate, no average position, and no query data. Google says clicks and position will arrive over time, with no date attached.

Why impressions on their own are an awkward gift

The report confirms that a page appeared in AI Overviews and stops there. It carries no click, no triggering query, and no position, so you are left knowing you showed up without knowing what the appearance earned. A page can pile up AI impressions and send nothing to your site, and the report cannot even surface that absence yet, since the click column does not exist. The signal is real but partial, so it has to be reasoned with rather than read straight off a dashboard.

One related control shipped the same day. Google added a toggle to remove your content from AI Overviews and AI Mode without losing position in regular Search. It exists, and it is the wrong lever for almost everyone, since opting out trades away the AI visibility this whole report was built to measure.

Pipe Search Console into an AI assistant

Since the data is thin, the move is to make it fast to interrogate. Connect Search Console to an AI assistant through an MCP server and the assistant can pull your standard Performance report live and reason over it in plain language, instead of you exporting CSVs and squinting at them.

Amin Foroutan's open-source mcp-gsc does exactly this, and it works with Claude and other MCP clients. There are two ways to run it. The hosted version offers one-click sign-in and needs no Python, while the self-hosted setup runs on Python 3.11 with your own Google OAuth. Credit to Amin for building and maintaining it.

One caveat decides how you run this. The standard Performance report comes straight through the connection, so the assistant queries it for you, no export needed. The Generative AI report does not, at least not yet, since Google has not confirmed it for the Search Console API. So that is the one report you export from the UI and drop into the same session. The assistant pulls the Performance data itself and reads your AI export beside it, which is the combination the prompt below relies on.

The proxy-query problem, and how to stay honest about it

The Generative AI report has no queries at all. To reason about intent, you borrow the page's standard web-search queries as a proxy for what might be triggering its AI appearances. That crutch is useful and easy to abuse. Google does not expose AI-specific queries, so those web queries are an approximation and never the actual prompts. Label them as a proxy in the analysis, and do not let anyone in the room present them as "AI prompts." The day Google does expose real AI queries, you want to have been the one who never overstated them.

A prompt that turns both reports into an analysis

Give the assistant the brief below. It pulls the Performance report over the connection, reads the Generative AI export you added to the session, and respects the gaps in the data instead of papering over them.

You are an SEO and GEO analyst with the Google Search Console MCP connection
live. Work on yourdomain.com for the last 30 days, from two inputs:
1. The standard Search performance report. Pull this yourself over the
connection: queries, pages, clicks, impressions, position.
2. The Generative AI report, attached as an export because it is not in the API
yet: impressions and pages in AI Overviews and AI Mode. It has no clicks and
no queries.
Treat the standard web-search queries as a proxy for intent, not as AI prompts.
Google does not expose AI-specific queries.
Analyze:
1. Pages with heavy AI impressions but few or no clicks. Use the standard report
for clicks and quantify the gap.
2. Whether each page's top query matches its topic, and where one broad term is
inflating impressions.
3. Branded vs non-branded, and any language or market splits such as /en vs /id:
where AI visibility concentrates, and where the clicks actually come from.
4. The top 10 prioritized fixes across content, schema, titles, internal links,
and GEO formatting. Separate quick wins from larger work.
Start with a five-bullet executive summary. Use tables with real numbers. Flag
any misleading metric.

What the framework is really checking for

Each instruction in that prompt is aimed at a specific failure. Pages with heavy AI impressions and thin clicks are where Google is using your content to answer a question without sending you the visit, and quantifying that gap tells you whether AI exposure is earning anything. Checking the top query against the page topic catches the common case where one broad head term inflates a page's numbers while contributing nothing qualified. The branded and language splits show where your visibility is concentrated against where revenue actually enters, which on a multi-market or multilingual site is often two different places. And forcing quick wins apart from larger work stops the output from becoming a wish list nobody actions.

What "GEO formatting" means in that fix list

GEO formatting is the bucket people gloss over, so it is worth naming. It means structuring a page so an AI feature can lift a clean answer from it. That means the answer stated plainly near the top, claims written in text rather than locked in an image or drawn by a script, structured data that names the entity, and a consistent identity across the page. AI surfaces pull from content a machine can extract, and a page that buries its answer or renders it in JavaScript can earn impressions and still never get cited cleanly. We have written separately about why client-rendered pages read as blank to most AI crawlers, and the same principle governs whether a page that appears in an AI Overview is one the model can actually quote.

The metrics that will mislead you

Three traps sit inside this report. AI impressions are not visits, and a page can appear in an AI Overview as one link in a citation row nobody opened. A broad head term can inflate a page's impressions while adding no qualified demand, which makes raw impression counts a poor success metric. And until clicks land in the report, any claim of "AI performance" built on impressions alone is half a measurement wearing the costume of a full one.

Where this leaves you

The Generative AI report is a genuine signal and a narrow one. It tells you where Google's own AI surfaces show your pages, and stays silent on ChatGPT, Gemini, and Perplexity, and on whether any of those answers describe you accurately. Treat it as one input among several.

Measuring AI visibility properly spans the assistants people actually ask and tracks four things the report cannot reach. It watches how often each engine mentions you, your citation share of voice against competitors, the sentiment of those mentions, and the traffic AI sends you. That is the work behind our AI Visibility Measurement, and the framework is published openly on our methodology page. Talk to us if you want the full picture, not just the slice Search Console can see.