How to interpret Google AI impressions in Search Console

What the Generative AI report counts

Start with the plain definition. Google's Search Console help page on impressions, position and clicks says an impression is recorded when someone has seen, or potentially seen, a link to your site in Search, Discover or News. A Google AI impression is that same event inside a generative AI feature: an AI Overview or AI Mode.

AI Overviews and AI Mode data sits inside the Web search type of Search Console's main Performance report, mixed in with every other result, and Google's AI features documentation still says so. On 3 June 2026, Google's Search Central blog announced a dedicated view alongside it: a separate Generative AI performance report. The post says the data will continue to be tracked in the main report too, and a note on it says the new report reached all websites worldwide on 31 August 2026. There are two versions: one for Search, and one for generative AI features in Discover.

Know what's missing before you read anything. There are no clicks, no click-through rate, no positions and no search queries. Google's launch post says it may add metrics over time. Clicks on links in AI Overviews and AI Mode still count in the main Performance report under Web, blended with ordinary results, and Google's documentation offers no filter to split them out. Search Console leaves out Search Labs experiments entirely.

What it does have, according to Google's help page for the Search version, is impressions: how many times links to your site were shown in a generative AI feature. It draws on the Web search type only, which you can split into text-based searches and multimodal ones, where an image was part of the search (Google names Google Lens and Circle to Search as examples). You can group the impressions four ways:

• Pages: the final URL the feature linked to, after redirects, with most data credited to the canonical URL Google chose.
• Countries: where the search came from. For an Australian business, filter to Australia first, since overseas searches rarely turn into local work.
• Devices: desktop, tablet or mobile (the Discover version has no device split).
• Dates: by day, week or month, in Pacific Time. The 24-hour view has no date table, and its chart uses your local time.

How a single AI impression is counted

An impression means a link was shown, or potentially shown. Nobody has to read it. Google's help page on impressions says standard impression rules apply to both AI Overviews and AI Mode. Under those rules, a link generally counts once it appears on the current page of results, even if the person never scrolls down to it. Links tucked inside an expanding or separately scrolling section typically have to be expanded or scrolled into view first.

A common worry is double counting when one page is cited in an AI Overview and also listed in the ordinary results for the same search. In the main Performance report, Google's About the data page says that when results are counted by property, a property that appears twice on one results page counts as a single impression, and the topmost position is the one reported. Counted by page, each unique URL is counted separately, so one URL shown twice for one search still reads as one impression for that URL, at its topmost position. The Generative AI report then counts the AI Overview appearance on its own.

That topmost-position rule needs care. An AI Overview holds one position on the results page, so a page cited in an Overview near the top can lift its average position in the main report without one extra visit. Check for that before you read a jump in position as a ranking win. For how these features choose which pages to show in the first place, see AI Overviews and AI Mode for local businesses.

These are the AI-specific rules from Google's impressions help page, and they apply in the main Performance report as well:

• An AI Overview takes a single position on the results page, and every link inside it shares that position.
• Clicking a link to an external page in an AI Overview or in AI Mode counts as a click. A tap that starts a new search or opens AI Mode, rather than opening a web page, isn't a click on an external page, so no site gets a click from it, even though your link may have been counted as shown.
• Position in AI Mode follows the same method as a normal results page.
• A follow-up question in AI Mode is treated as a new query, so its impressions and clicks belong to that follow-up. One conversation can therefore produce several impressions for the same page.

Why the chart and table totals differ

Expect them not to match. Google's help page for the report says the chart is aggregated by property: if two results from your site appear in the same generative AI feature, the chart counts one impression. Group the table by page and each URL is counted on its own, so the page rows can add up to more than the chart total. Add a URL filter and the chart switches to counting by URL.

Other limits carry over from the main Performance report. Tables show at most 1,000 rows, so on a site with many pages, rare page rows can drop out of the table while still counting in the chart. Outside the 24-hour view, days are labelled in Pacific Time, which puts the report's day boundaries out of step with an Australian business day. The newest data can be preliminary, drawn as a dotted line, and may change within hours.

Quote the chart total when the question is how often your site appeared in AI features at all. Quote the page table when the question is which pages did the appearing. Never add the page rows together and call the result your total. The table below sets out each view.

Blanks in the report become zeros in your export

This one quietly breaks spreadsheets. Google's help page for the report says values shown as ~ or - (not available, or not a number) are written as zeros in the downloaded file. Once exported, a blank on screen and a real zero look identical.

You'll likely be working from those files, too. Google's Search Console API reference, last updated in August 2026, lists web, image, video, news, Discover and Google News as data types, with no separate generative AI type. For now, the export button is the documented way to get this data out.

The main Performance report has the same export rule, and it bites hardest on position. Google's impressions help page explains that a dash in the position column means no position was recorded, because nobody saw your site for that query. Averaged as a zero, a missing position makes your average look better than it is. Before you chart an export, work through these steps:

1. Open the same view on screen and note every cell showing ~ or -.
2. In the file, replace the zeros in those cells with a blank or 'not available'.
3. Write down the dimension and filters you used, because the aggregation changes with them.
4. Label the dates as Pacific Time before you join them to Google Analytics, call logs or booking data.
5. Use weekly or monthly granularity for trends, which Google says smooths out daily swings such as weekends.
6. Keep the raw download untouched and make your fixes in a copy, named with the date you exported it, because preliminary days can change after you download.

Reading a rise or a drop in AI impressions

Rule out reporting causes first, then compare the AI trend with the same page's Web clicks over the same weeks.

Start with Google's data anomalies page in Search Console Help, which records known issues that might affect your data. It lists a logging error that caused a decrease in impressions on the Generative AI performance report in Search for data from 13 to 17 August 2026, and an update dated 21 August says the missing data has been restored and the Performance report now shows complete metrics. So that entry only explains a mid-August dip in an export or screenshot taken before 21 August: export the period again and replace the old figures. If the dip still shows in the live report, the logging error is not the cause, so work through the other checks below.

Rising AI impressions with flat or falling Web clicks is the pattern people ask about most. Read it in this order. Filter both reports to the same page, the same country and the same weeks, at weekly granularity. Check whether the page's average position in the main report jumped, since an AI Overview near the top can move it. Check whether searches for the topic usually rise or fall at this time of year (see local SEO seasonality and reporting). Only then compare clicks and click-through rate. If position and demand held steady while AI impressions rose and clicks fell, the pattern fits people getting what they need from the AI answer. The report can't prove that, because it holds no clicks and the main report blends AI clicks with ordinary ones.

Use this table to match what you see with the first thing to check. No one can make Google show an AI feature, or your link inside one, so treat every row as a diagnosis rather than a fix:

A worked example: a clinic's month-end check

Say a physiotherapy clinic in Maroochydore opens the report for the first time, filtered to Australia. This business is hypothetical, and so is every figure below. The chart shows a modest total. Grouped by page, the rows add up to noticeably more, because its treatment page and its booking page sometimes appear together in one AI Overview, and the table counts each. In the report for the practice manager, the owner quotes the chart figure as the site's AI impressions and lists the page rows underneath, labelled as per-page counts that overlap.

Next, the owner compares this month with last month, still grouped by page, and exports the table for the practice manager. A newer page on running injuries was rare last month, so on screen its last-month cell showed ~. Google's help page on filtering and comparison, which the report's help page links to, says a value that is very rare in one compared group but not the other shows ~ for the rare group. In the file, that cell is a 0, which reads as a page that was never shown. The owner marks it as not available, then filters the report to that one page, where Google says values for both periods should appear. Last, they switch the chart to weekly granularity, which the same help page suggests when comparing two date ranges, and export again for the trend.

Then they size it. Both reports are set to the same week, Australia and weekly granularity, and the owner reads the two chart totals, which both count by property. Say, purely as an illustration, the Generative AI chart shows 300 impressions and the main report's Web chart shows 4,000. Dividing one by the other gives 7.5 per cent: a rough guide to how often the clinic's appearances involved an AI feature that week. It stays rough. The AI figure already sits inside the Web figure, and a results page that showed the clinic in both an AI Overview and an ordinary result counts once in the Web chart. Using a page table instead of the chart would mix two kinds of counting and give a meaningless share.

The report has no queries, so the owner looks for clues in the main Performance report. They filter Pages to the running injuries URL, then add a Queries filter using Custom (regex), which Google's filtering help page says uses RE2 syntax. A pattern such as ^(how|what|why|can|is|do)\s keeps question-style searches. These are searches that showed the page in any kind of result, so a question-shaped query is a hint about the topics behind the AI impressions, and nothing more. Search length or wording can't identify an AI result, and Google leaves rare anonymised queries out of the table.

Now the data points to work. The treatment page collects AI impressions but few Web clicks, so the owner checks that it answers its main question in the first lines and gives a clear reason to visit, such as booking times. The clinic's core service page for back pain has no AI impressions at all, so they confirm it's indexed and that it answers the questions patients actually ask (see how to rank in Google AI Overviews). A page that dropped out last month gets a redirect and canonical check before anything is rewritten.

The clinic can now say which pages Google shows in AI features and whether that's rising. It still can't say those impressions produced a single booking. That answer lives in its enquiry records, which is why the numbers need matching against call tracking and conversion reconciliation.

What the numbers can and can't tell you

If the report is missing or flat, check two things. Google's help page says a missing report can mean the site hasn't had enough impressions in these features yet. And Search Console has a Search generative AI control under Settings. Google's page on that control says include is the default, but a property set to exclude, or one inheriting that choice from a parent property, won't appear in AI Overviews, AI Mode or Discover's generative AI features, so there are no impressions to count.

Bing reports a different unit. Microsoft's Bing Webmaster blog post of 10 February 2026 introduced an AI Performance report, in public preview, covering Microsoft Copilot, AI-generated summaries in Bing and select partner integrations. It counts citations, the times your pages were displayed as sources in AI answers, and it shows grounding queries, which Google's report doesn't. Microsoft says the data doesn't indicate ranking, authority or a page's role within an answer. A Google impression and a Bing citation come from different products with different counting rules, so set them side by side as two trends and never add them together. If you also pay for a third-party AI tracking tool, compare its method with Search Console's first.

Google's AI features documentation suggests tracking conversions and time on site in a tool such as Google Analytics alongside Search Console. When AI impressions go into a monthly SEO report, keep the entry short and honest. A useful format covers five things:

• The trend: weekly AI impressions from the chart total, filtered to Australia, over at least the last three months.
• The pages: the top pages by AI impressions, plus any core service page with none.
• The changes: each rise or drop with the check that explained it, such as an anomaly date, a redirect or a seasonal pattern.
• The limits: no clicks, no queries, Pacific Time dates and corrected export zeros, stated in one line.
• The business result, taken from call tracking, forms and bookings, never inferred from impressions.