Generative AI in Search Console

#Where it is and what it shows

The section lives in the Performance report, under the Search results view, as a separate Generative AI entry. The data breaks down by pages, countries, devices and dates, down to hourly granularity. History goes back to May 18, 2026 at most, because Google did not collect it in this form before then.

The rollout is gradual. If you do not see the section on your account yet, that is normal. Selected domains got it first, and the rest are arriving in batches.

#AI Overviews and AI Mode are two different things

The report merges two surfaces that are worth keeping separate in your head. AI Overviews are the summaries generated above classic results, which the user sees without asking for anything beyond a normal query. It is a passive surface. AI Mode is a separate conversational search mode, where the user deliberately asks questions as a dialogue. That is active intent, and different behaviour. The Generative AI section counts both together, with no split between the two surfaces.

For anyone reading the report, that has a practical effect: you cannot tell today whether your impressions come from passive summaries or active conversations. Those are two different paths to your content, and likely two different audiences. Google does not separate them yet, so keep it in mind when you interpret the numbers.

#What the report leaves out

This is the part that matters most. The report shows impressions only. There are no:

• clicks,
• CTR,
• average position,
• queries.

So you learn which pages are cited in AI answers, but not whether anyone clicked through. For judging return on investment, that is not enough. The UK regulator CMA has required Google to provide clicks and CTR within nine months, so a fuller picture should follow. For now you have half the answer.

#The key caveat: this is not new data

It is easy to assume Google has revealed a new traffic source. It has not. These impressions were already counted in the standard Performance report. The Generative AI section is a filtered view of the same data, just better organised.

That has a practical consequence. If you see thousands of impressions in this section, do not add them to your existing traffic. They were already there.

#How an impression is counted

Two things are worth remembering:

• A link has to be scrolled or expanded into view, not just present in the background of an answer. Being among sources the user never expanded does not count.
• Aggregation per service. If two URLs from your site land in one AI answer, that counts as a single impression.

Without that context it is easy to over-read or under-read your presence in AI.

#What each dimension tells you

With so little data, it is worth squeezing what each dimension carries:

• Pages. The most important one. It shows which specific pages AI picks from your site as a source. That is a list of content that already earned a citation, and a hint about what to expand so more of it gets picked.
• Countries. Tells you which markets you appear in inside AI answers. If you run a multilingual site, it is a check on whether your AI visibility matches the distribution you actually care about.
• Devices. The desktop versus mobile split shows where your AI presence is stronger. It is often different from your classic results.
• Dates. Hourly granularity lets you see whether a publish or an update translated into a generative presence, and after how long.

None of these dimensions tells you whether anyone clicked. But put together, they map what AI takes from your site, where, and when.

#Getting the numbers out of the interface

The Performance report caps a UI export at 1,000 rows, which is fine for a small site and useless for one with thousands of URLs. Two routes take you past that.

The first is the Search Analytics API, a POST to https://www.googleapis.com/webmasters/v3/sites/{siteUrl}/searchAnalytics/query. The searchAppearance dimension is the one that carries feature types, and it has a rule that trips people up: you have to request it on its own, without other dimensions, then use the values it returns as a dimensionFilterGroups filter on a second call. So the workflow is always two requests, not one.

# 1. discover which appearance types exist on this property
curl -s -X POST \
  "https://www.googleapis.com/webmasters/v3/sites/sc-domain%3Aexample.com/searchAnalytics/query" \
  -H "Authorization: Bearer $TOKEN" -H "Content-Type: application/json" \
  -d '{"startDate":"2026-05-18","endDate":"2026-07-01","dimensions":["searchAppearance"],"dataState":"all"}' | jq '.rows[].keys'

Do not assume an appearance value before you have seen it in that first response. Google adds and renames these over time, and hard-coding a string you read in a blog post is how an export silently returns zero rows. Set rowLimit to 25000 (the per-request maximum) and page with startRow. Add "dataState":"all" if you want the most recent, still-incomplete days included, and leave it out when you are building a stable monthly baseline.

The second route is the bulk data export to BigQuery, configured in Settings. It writes searchdata_site_impression and searchdata_url_impression daily. That pair is also the cleanest illustration of the aggregation rule described above: the site table counts one impression per property per result, the URL table counts per URL. Before writing any query against a generative feature flag, inspect the schema, because the boolean feature columns change as Google ships new surfaces:

SELECT column_name, data_type
FROM `your_project.searchconsole.INFORMATION_SCHEMA.COLUMNS`
WHERE table_name = 'searchdata_url_impression'
ORDER BY column_name;

If the flag you need is not in that list yet, the API is the only source, and Looker Studio connected to the same property is the cheapest way to keep a chart of it.

#Checks to run before you trust a change in the graph

A moving line in this report has more boring explanations than interesting ones. Work through them in order.

Check the property type. A domain property aggregates every subdomain and both protocols. A URL-prefix property does not. If you compare a Generative AI number from one against a total from the other, the share you calculate is wrong before you start.

Check the date range against the data floor. Nothing exists before May 18, 2026. Any month-over-month comparison that reaches back further is comparing a real number against an absence, not a decline.

Check whether the last days are still filling in. The most recent days in the chart keep updating for a while after the fact. Read a trend from a range that ends several days ago, not from the right edge.

Check the time zone. Search Console reports in Pacific time, while your analytics tool almost certainly reports in your own. A daily comparison between the two will look off by a slice of a day, permanently, and no amount of rechecking the numbers will fix it.

Check the filters you left on. Country, device and page filters persist across views in the interface more often than people expect. If a number looks implausible, clear every filter and read the unfiltered total before you investigate anything else.

#Common mistakes reading this report

Three slip-ups come up most often:

1. Adding AI impressions to your existing traffic. They were already counted. Doing that is double counting the same data.
2. Confusing an impression with a visit. An impression means the link was visible in the answer, not that anyone came to the page. Without clicks, you do not know how many people actually reached you.
3. Treating missing data as a decline. If you do not have the section yet, or the history only goes back to May 18, that does not mean you were not cited before. It means Google did not report it before.

#Failure modes that look like a drop

Once you have a few weeks of history, the first alarming dip will arrive. These are the causes worth ruling out before you touch the content.

A URL change without a redirect. The pages dimension is keyed on the exact URL. Change a slug, and the old URL stops accumulating while the new one starts from zero. The property total barely moves, but any page-level dashboard you built shows a cliff. Before concluding anything, request the old path and confirm it still returns a 301 to the new one, wherever your redirects live: a server config, a host-level redirects file, or a plugin.

Canonical consolidation. Impressions are attributed to the canonical URL, not the one the user could have reached. If a canonical moved from a long slug to a short one, the reporting moved with it. That is a bookkeeping change, not a visibility change.

A query that never triggered a generative answer. Google does not fire an AI Overview on every query, and the trigger rate for a given intent is not stable. A page can hold its classic position perfectly while its generative impressions go to zero, simply because the surface stopped appearing for those queries. Compare against the classic Performance numbers for the same page and date range before assuming the page lost anything.

Seasonality in the underlying demand. Generative impressions inherit the seasonality of the queries behind them. A B2B page will dip across holidays in its main market and recover, and the countries dimension will show you exactly which market caused it.

A snippet directive someone added. A nosnippet tag or a tightened max-snippet value shipped in an unrelated release will zero a page’s generative impressions while its classic position holds steady. It is the only cause on this list you can reverse the same day, and the next section covers how to detect it.

#What actually controls whether a page can be cited

There is no separate switch for AI Overviews inside Google Search. Inclusion runs on the same snippet controls that have governed rich results for years, which means an ordinary robots meta tag can remove a page from generative answers while it still ranks.

• <meta name="robots" content="nosnippet"> removes the text snippet, and with it the ability to be used in a generated answer.
• max-snippet:[n] caps the characters Google may use. A tight cap set years ago for a different reason still applies today.
• data-nosnippet on a div, span or section excludes just that block. Applied to the paragraph that answers the question, it removes exactly the part that would have been cited.
• Disallow in robots.txt blocks the fetch entirely, so nothing can be quoted regardless of the meta tags.

Google-Extended is a different lever. It governs Gemini apps and grounded generation in Google’s AI products, not inclusion in Search AI Overviews. Blocking it does not remove you from the Generative AI report, and allowing it does not add you. Teams conflate the two constantly, then read the report looking for an effect that was never wired up.

Before you attribute a drop to an algorithm, run the mechanical check on the affected URL:

curl -sI https://example.com/en/your-page/ | grep -i x-robots-tag
curl -s https://example.com/en/your-page/ | grep -Eo '<meta[^>]+name="robots"[^>]*>'
curl -s https://example.com/en/your-page/ | grep -c 'data-nosnippet'

An X-Robots-Tag header set at the CDN or server level is the one people forget, because it is invisible in the page source and in the CMS. A CDN transform rule, a reverse proxy or a security plugin can all set it without touching the markup, which is why reading the response headers is the only reliable check.

#What to do with it in practice

Since all you have is impressions, get what you can from them:

1. Work out the AI share of your impressions. Compare the Generative AI section against the total in the Performance report, over an identical date range and with an identical set of filters, or the ratio means nothing. Reported shares swing hard with market and query mix, so there is no industry benchmark worth borrowing here; the only figure that applies to you is the one you calculate. A worked example: 128,000 total impressions in a month against 21,000 in the Generative AI section puts your AI share at about 16 percent. Record that number, because its change over time says more than the value itself.
2. See which pages get cited. The pages dimension shows what AI picks from your site. That is a real signal for which content to expand and which to rewrite so it gets cited more often.
3. Do not optimise for clicks you cannot see yet. Until Google adds CTR, treat impressions as a presence signal, not a conversion one.

#Building the baseline before clicks arrive

The value of this report today is almost entirely in the archive you start keeping now. When clicks land, everyone will have the same fresh data and nobody will have history, unless they stored it.

Keep it small enough that it survives contact with a busy month. One row per property, per month, per page group, with five fields: date range, total impressions, generative impressions, generative share, and a free-text note for anything that happened (a migration, a redesign, a content pass). A CSV in the repo or a single BigQuery table is enough. A dashboard that nobody rebuilds after it breaks is not.

Two habits make the archive usable later:

Group pages before you store them. Page-level rows are noisy and turn over with every slug change. Grouping by section (blog, service pillars, city pages) keeps the series comparable across a year and survives the URL churn described above.

Log the interventions on the same timeline. A generative share that moved from one month to the next is uninterpretable without knowing what you shipped. One line per change, dated, in the same file. When CTR arrives, the question will be which content shape converts from an AI answer, and that is only answerable if you can line up changes against the series.

The first analysis worth running the day clicks appear is the gap between classic CTR and generative CTR for the same pages. A page that holds both is genuinely earning traffic from the surface. A page with high generative impressions and near-zero clicks is being summarised, and the value it produces is brand exposure, not sessions. Those two cases deserve opposite treatment, and today you cannot tell them apart.

#What changes when Google adds clicks

Today the report answers one question: which pages are popular in generative features. Clicks and CTR will change that into a far more important one: whether AI presence turns into traffic, or just lets the model summarise your content and keep the user on its own surface. That distinction decides whether AI visibility is an acquisition channel or a cost.

The timing is not accidental. The nine-month CMA window is what puts clicks on the roadmap at all, so treat the delivery date as a regulatory deadline rather than a product promise, and treat the archive described in the previous section as the thing that will make the first CTR figures readable. Without that base, the first CTR data will be a number with no context.

The report is a measurement tool, not a strategy. If you want AI to cite your pages more often, that is content and structured-data work, which we cover in the GEO guide. If you want it implemented and measured on your site, that is our GEO/LLMO optimization service.