You open Search Console on a Tuesday, click through to Search results, and there is a filter option that was not there a few quarters ago: AI Mode, sitting alongside AI Overviews in the search type menu. Clicks look flat. Impressions look odd. Somebody in the business has already asked whether AI is eating your traffic. Most guides to the search console ai mode report stop at explaining what Google launched, which is the least useful part. What follows is a working routine instead: four short passes, fifteen minutes, ending with three written decisions and a note of the numbers you should stop chasing because Google does not publish them.

What Google actually added, and where to find it

The route in is unglamorous. Performance, then Search results, then the search type filter at the top of the chart. Alongside the classic Web, Image, Video and News types, Google now exposes generative AI views: AI Overviews, AI Mode, and generative AI features appearing in Discover. That is the whole navigation story.

The part that causes the first misreading is arithmetic. AI Mode clicks and impressions are not a separate channel bolted onto your totals: they sit inside the Web search type figure. If you add the AI Mode numbers to your Web numbers and report the sum, you have double counted, and someone will eventually notice. Treat the generative view as a slice of a pie you already own, not a new pie.

History is short, and that is normal. Google announced generative AI performance reporting in Search Console in June 2026 and extended it worldwide by late August 2026, so a UK site opening the report now has months of data, not years. The absence of a 2024 baseline is not a tracking fault, a tagging error, or something your developer broke. It is the rollout. Google's own Search Console performance report documentation is the place to confirm what is currently available, because the feature set has moved more than once.

One habit before you interpret anything: export first. Download the page table and the query table to a spreadsheet with the date in the filename. Search Console keeps 16 months and then quietly drops the oldest data, and the comparison you will want in eighteen months is the one you did not save.

Minutes 0 to 3: set the comparison before you read a single number

Nearly every bad conclusion from this report starts with a lazy date range. Month-to-date is the worst offender: on the 6th of the month it makes you look destroyed, and on the 29th it flatters you. Pick 28 days against the previous 28 days. Equal weekday counts, equal weekend counts, no calendar accidents.

Then filter to the United Kingdom if that is where your customers are. Overseas impressions drag position averages around with no commercial meaning attached, and a service business in Leeds does not need Indonesian AI Mode impressions in its baseline.

Switch all four metrics on: clicks, impressions, CTR and average position. Clicks alone tell you almost nothing on generative surfaces, because the interesting movement happens in the gap between impressions and clicks. Then write your four baseline figures down somewhere outside Search Console. The interface does not keep a reading history for you, and next month you will not remember whether 1,400 impressions was up or down.

Minutes 3 to 8: the page pass that shows which content AI is using

Sort by impressions. Not clicks. This is the single change that makes the report useful, because impressions move on AI surfaces long before clicks do: a page can be feeding answers for weeks while its click line stays perfectly boring. Sort by clicks and you see your usual winners. Sort by impressions and you see what Google is actually pulling from.

Read the relationship per page, not the totals. Two patterns matter.

• AI impressions up, clicks flat or down. Almost always definitional or how-to content: what is X, how does Y work, step-by-step guides with nothing in them that a summary cannot carry.

• AI impressions up, clicks holding. Look at what those pages contain. Real prices. A named process with stages. Dated examples. First-hand specifics, screenshots, a number somebody had to actually find out.

That second pattern is a content instruction, not a coincidence. A generative answer can paraphrase a definition in one line. It cannot carry your installation lead times, your actual quote ranges, the three things that go wrong on site, or the name of the process you run. Pages that survive summarisation are pages that contain something specific enough to be worth clicking through for.

Finish the pass by cross-checking against your publishing log. If a page jumped, did you ship something, refresh something, or get linked from somewhere? Movement you can attribute is movement you can repeat. Movement you cannot attribute is usually Google.

Minutes 8 to 12: queries, devices and countries, and the limits built into each

The query table will never add up to your totals. Google withholds low-volume and personal queries for privacy, so the sum of your query rows always falls short of the chart above them. Chasing that gap is the biggest time-waster in this entire report. We have watched teams spend an afternoon on it. Treat queries as a sample for spotting themes: which topics, which intents, which phrasings, and nothing more precise than that.

Average position deserves a similar warning. Position inside generative features is calculated differently from a classic blue-link ranking, and a summary with three cited sources is not a top ten in the sense you are used to. Read it as directional. Anchor your decisions to impressions and clicks per page, and keep proper rank tracking as a separate cross-check rather than a replacement.

Give devices thirty seconds. Generative surfaces behave differently on mobile, where the answer occupies more of the screen before anything else appears, so a page that is mostly mobile traffic is a different refresh priority from one that is mostly desktop. That is the whole analysis. Do not build a strategy on it.

What the report cannot tell you is worth saying plainly, because people keep looking for it: whether you were cited in a given answer, where your link sat inside it, and which competitors were cited beside you. None of that is in Search Console. If you want that picture, you look at the answers themselves.

Minutes 12 to 15: turn the reading into three decisions

Fifteen minutes is enough for exactly three decisions. Any more and you are redesigning your content plan on a dashboard, which is how quarters get wasted.

Decision one: one page to refresh. Pick the clearest case of AI impressions up with clicks down, and add what a summary cannot replace. A service page in that position does not need a rewrite. It needs pricing ranges with a date on them, the named steps of how you actually deliver the work, what is excluded, and how long each stage takes. We would take that over another 1,200 words of introduction every time.

Decision two: one thing to leave alone. Most single-week swings are noise, especially in a report this young. If a page is moving inside its normal variance, write down that you are deliberately not touching it. Recording a non-decision stops you touching it next week and then wondering what caused the change.

Decision three: one article to plan. Find a topic picking up impressions where your coverage is thin, and brief one piece. One. Not three near-identical explainers, which is the reflex reaction and a reliable way to cannibalise the page you already rank with.

Then write the three lines down with today's date and the figure that triggered each one. Something as plain as: "6 November 2026, refresh /commercial-cleaning, AI impressions clearly up vs prior 28 days, clicks flat, adding price bands and process. Leaving /blog/what-is-soft-washing alone, within variance. Planning one piece on contract cleaning SLAs, impressions appearing, no page exists."Next month that note is your comparison, and it is better than memory.

Four misreadings that cost teams a quarter of wasted work

1. Reading a drop as a penalty. AI impressions fall when the query mix shifts or when Google changes what triggers a generative answer. That is a feature decision on Google's side, not a manual action against you. If you genuinely suspect something algorithmic, our notes on checking an unconfirmed Google update cover what to look at first.

2. Rebuilding strategy on four weeks of data. The report is new, the surfaces behind it are changing, and four weeks is a sample, not a trend.

3. Answering AI summaries with more shallow explainers. If a summary replaced your definition page, publishing two more definition pages hands Google more material to summarise. Go the other way: deeper, more specific, harder to paraphrase.

4. Reporting AI Mode to the board as a channel. It is a slice of existing search data. Present it inside your search reporting, with the double-counting caveat said out loud, or you will be explaining the discrepancy for the next three meetings.

Where this fits in a monthly content routine

Monthly beats daily for most SMEs, and it is not close. A report with months of history cannot support weekly conclusions, and content changes take weeks to show up anyway. Weekly only makes sense if you publish weekly. Between readings, stop opening it.

Pair the fifteen minutes with two other things: a glance at your rank tracker for the handful of terms that actually convert, and a page-level log of what you published, when, and what moved afterwards. One dashboard is never enough to judge visibility, and the log is what turns a reading into evidence.

The editorial half matters more than the measurement half. We plan briefs against the live top ten rather than against a keyword tool's guess, and we treat refreshing an existing article as a higher-value job than adding a new one when the data says a page is being summarised rather than clicked. That is the method behind how we write and research every piece, and it is the same logic that keeps content inside Google's guidelines for AI-assisted publishing: original, specific, written for a reader who wants the detail. If you would rather see the whole pipeline before you commit to anything, the free content marketing snapshot is a reasonable starting point.

Set the cadence, keep the exports, and let the search console ai mode report do one job: telling you which pages to make more specific. It will not tell you who cited you, and it will never reconcile to the last impression. Fifteen minutes and three written decisions is the correct amount of attention to give it.

Frequently Asked Questions

Where is the AI Mode report in Google Search Console?

Open Search Console, go to Performance, then Search results, then the search type filter above the chart. Alongside Web, Image, Video and News you will find the generative AI views, including AI Overviews and AI Mode. Availability has changed more than once since launch, so check Google's current performance report documentation if a view is missing for your property.

Are AI Mode clicks and impressions counted separately from Web search in Search Console?

No. They are included within the Web search type totals, which is why adding them on top produces an inflated number. Use the generative view to see the slice, and report the Web total as your headline search figure. Say the overlap out loud in any board report, because somebody will check the sums.

Why do the queries in the generative AI performance report not add up to my totals?

Google withholds queries that are rare or that could identify an individual, so the query table is always a partial sample of the traffic shown in the chart. The gap is deliberate and permanent, and no amount of filtering will close it. Use queries to spot themes and intents, and use the page table for anything you intend to act on.

How often should I check the Search Console AI Mode report?

Monthly suits most small and medium businesses. The data history is short, weekly swings carry very little signal, and any content change you make needs several weeks before its effect is readable. Move to weekly only if you are publishing weekly and have enough volume for the numbers to settle.

Does a drop in AI Mode impressions mean my site has been penalised?

Usually not. Impressions on generative surfaces fall when Google changes what triggers an AI answer, or when the mix of queries people type shifts, neither of which is a judgement on your site. Check the Manual Actions and Security Issues panels in Search Console if you want certainty, then look at whether your classic Web clicks moved at the same time.