AI Mode SEO: What 24 Million Impressions Across Six Sites Were Actually Worth

Across six Search Console properties over 92 days, 24,498,104 impressions returned 210,102 clicks. That is a blended click-through rate of 0.86%. One property in the set, the one this study then takes apart query by query, produced 57,968 impressions and 93 clicks over the same window. That is 0.16%. Google now tells you how often you appear in AI Overviews and AI Mode. It does not tell you what appearing is worth. This is an attempt to put a number on it.

Key takeaways

• Across six properties and 24.5 million impressions over 92 days, blended search CTR was 0.86%. Per-site rates ran from 0.09% to 0.91%, a tenfold spread.
• On the anchor property, 37 queries held an average position of ten or better with at least ten impressions each. Between them they earned zero clicks.
• Every single click that property earned from a top-ten average position came from a query with six impressions or fewer.
• 99 non-branded question-form queries carried 5,597 impressions and returned no clicks at all.
• Impressions rose 13-fold across two consecutive 46-day windows while average position fell from 17.7 to 31.8. Both numbers were correct, and read together they mean the opposite of what they appear to mean.

The number Google gave you, and the four it kept

Search Console’s generative AI performance report is the first official window into AI visibility. Google’s Search Console documentation states that it covers impressions from AI Overviews and AI Mode. Per Google’s own report documentation, you can group that data four ways, by Pages, Countries, Dates and Devices, and impressions is the only metric on offer.

Now the part that matters. According to Sinton Agency’s August 2026 breakdown, the report leaves out queries, clicks, click-through rate, average position, where in the answer your citation appeared, which passage of your page was used, and anything about conversions or revenue.

Read that list again. Google shipped a visibility metric with the entire value half removed.

This is a genuinely awkward position for anyone who reports to a client or a board. You can now say “we appeared 58,000 times in AI answers.” You cannot say whether that was worth anything, because the number that would tell you is the one Google kept. An impressions chart with no clicks attached is not a performance report. It is a presence report, and presence has never been the thing anyone was buying.

So we went and got the missing half from the only place it still exists.

What we measured, and how you can repeat it

Six Search Console properties. One 92-day window, 25 May to 24 August 2026. Every figure below was pulled live on 2 September 2026 and recomputed from scratch rather than carried over from an earlier draft, which turned out to matter.

The properties span six verticals: an SEO consultancy, executive coaching, a digital agency, consumer SaaS, a design studio and a developer tool. They range from 9,132 impressions to 21.9 million, which is deliberate. A study drawn only from large sites would tell you nothing about the small ones.

All six are anonymised at property level, and I am not naming any of them. The query-level detail in Findings 2 and 3 comes from one property, called Site A throughout and referred to below as the anchor property, where I hold both the access and the permission to publish query text. That permission is the only reason query-level analysis appears here at all.

The join: what stands in for the clicks Google withheld

There is no clever trick here, and I want to be precise about that because the method is the whole argument.

Google’s generative AI report gives impressions with no clicks. But Google’s documentation on the report’s data states that the generative AI report draws from the same Web search type that feeds the ordinary performance report. The AI impressions are not a separate pool. They are already inside your normal Search Console numbers, unlabelled, mixed in with everything else.

That is the join. You do not need to reconstruct AI clicks from a second system. You need to read total Search performance for a window in which AI surfaces were scaling, and understand that the AI impressions are in there, dragging the aggregates around. Every number in this study is total Web search performance from the Search Console API, pulled per property for the same 92 days.

What we could not control for

Stated here, before the findings, because findings should be spent against credibility already earned.

We cannot isolate AI-surface clicks. Nobody can, including Google, or the report would have shipped with them. What we can measure is what happened to total search performance while AI surfaces grew, which is a weaker claim than “an AI Mode impression is worth £X” and is the honest version of it.

Query-level data is incomplete by design. Google’s stated data limitations confirm the 1,000-row limit applies to this report as it does to the performance report. On the anchor property the query table accounts for 37,099 of 57,968 impressions, about 64%, but only 8 of 93 clicks. Anonymised queries hide most of the clicks. Every query-level finding below is therefore a statement about the visible 64%, not the whole.

Aggregation differs between views. Google on the report’s aggregation notes that country, device and date data is aggregated by property while page data is aggregated by page, and that chart and table totals can differ as a result. If your numbers do not reconcile, this is usually why.

And the window is not a controlled experiment. Sites published content, gained links and changed during it. This is observational.

Finding 1: what one impression is actually worth

Property	Vertical	Impressions	Clicks	CTR
Site A	SEO consultancy	57,968	93	0.16%
Site B	Executive coaching	702,264	6,171	0.88%
Site C	Digital agency	9,132	76	0.83%
Site D	Consumer SaaS	21,870,976	198,815	0.91%
Site E	Design studio	14,981	13	0.09%
Site F	Developer tool	1,842,783	4,934	0.27%
Blended	six properties	24,498,104	210,102	0.86%

The blended figure of 0.86% is dominated by the consumer SaaS property, which contributes 89% of all impressions in the study. The median site rate is 0.55%, and the honest summary of this table is the spread rather than the blend: 0.09% to 0.91%, a tenfold range across six sites in one quarter.

One impression is worth, on average across these six properties, 0.0086 of a click. Turn that around: you need roughly 117 impressions to buy one visit. On the design studio property you need 1,152.

The three properties at the bottom of that table share a trait. Each grew impressions sharply during the window while clicks stayed close to flat. The design studio earned 13 clicks across the whole quarter on 14,981 impressions, and fewer than 150 of those impressions landed before July. That is the shape of the thing. Impressions are being handed out generously. Clicks are not.

I would not report the blended number to a client. I would report their own rate and this range, so they can see where they sit in it.

Finding 2: question-shaped queries are where impressions go to die

This is the finding I would most want another analyst to try to break, because it is the one with the clearest instruction attached.

I took the 100 highest-impression question-form queries on the anchor property, meaning queries beginning how, what, why, when, where, who, which, can, do, does, is, are, should or will. Between them: 5,598 impressions and one click.

That one click came from a branded query on the site’s own name, carrying a single impression. Strip it out and you get the number I keep coming back to: 99 non-branded question-form queries, 5,597 impressions, zero clicks.

Position was not the problem. Eighteen of those queries held an average position of ten or better and carried 1,619 impressions between them. Still nothing. “How can I structure content to win more featured snippets?” sat at average position 7.0 across 650 impressions and returned not one visit. “How often should I run a geo audit” averaged position 2.4. Nothing.

Google’s definition of an impression is that a link to your site was shown to a user inside a generative AI feature. Shown. That is the whole guarantee. When someone asks a complete question and the surface answers it inline, being shown underneath that answer is the end of the transaction, not the start of one.

The practical consequence is uncomfortable if you have been building question-led content for AI visibility. Question-shaped headings do appear to earn the impression. On this evidence they do not earn the visit. If you are planning content against AI surfaces, the tracking setup worth having matters more than the heading format, because the heading format is optimising for a metric that terminates.

Finding 3: average position stops meaning what you think

Split the 92 days into two consecutive 46-day halves and the anchor property looks like this.

	Impressions	Clicks	CTR	Average position
25 May to 9 Jul	4,129	14	0.34%	17.7
10 Jul to 24 Aug	53,839	79	0.15%	31.8

Impressions rose thirteenfold. Average position got fourteen places worse. Both numbers are correct.

They move in opposite directions because they are not measuring the same thing any more. Google’s documentation on the report’s data puts generative AI impressions inside the Web search type, which is the same pool your average position is calculated over. When AI surfaces start showing your pages in deep positions they never used to reach, every one of those appearances enters the average as a bad number. Visibility rose. The metric that is supposed to describe visibility fell, because a flood of low-position impressions arrived and dragged the mean down.

On this property 81% of query-attributed impressions sat below position 20, and positions 21 and deeper produced exactly one click across the whole quarter on 30,054 impressions.

The trap is reading the two together and concluding you have been demoted. You have not. You have been shown more, worse. If you track average position as a KPI, it stopped being a clean signal the moment AI impressions entered the pool, and a declining line is now as likely to mean growth as decline.

Here is the sharpest version of it, and the number I would put in front of a client. On the anchor property, 37 queries held an average position of ten or better with at least ten impressions each, carrying 2,528 impressions between them. They earned zero clicks. Every click that property earned from a top-ten average position, four in total, came from a query with six impressions or fewer. Top-ten positions on queries anyone actually searches returned nothing at all.

How to report AI visibility without lying to yourself

Google’s rollout note, August 2026 confirms that as of 31 August the report reached all websites worldwide. So this stopped being a thing that happens to other people about a day before this was published. Whatever you report next month will include it.

Three things I would change in a client report now.

Retire average position as a headline. Not as a diagnostic, where it still earns its place per page and per query, but as a top-line trend line it has become actively misleading for the reason above. If you keep it, put impressions next to it and never show one without the other.

Report impressions and clicks as a pair, always, with the ratio. An impressions chart on its own is the exact failure mode Google’s report invites. The ratio is the number that carries information: it tells you whether growing visibility is converting into anything. On these six properties it mostly is not.

Add a line for what the visibility is not doing. If AI impressions tripled and clicks moved 4%, that belongs in the report as a finding, not buried as an absence. It is genuinely useful intelligence: it tells you the surface is answering the query itself, and that the content strategy pointed at that query needs rethinking rather than more of the same.

What I would not do is convert impressions into an estimated traffic value using a CTR curve built for blue links. The curves are wrong now. This study’s own numbers show top-ten positions returning zero, which no classical CTR model predicts. If you need a dashboard that holds both numbers honestly, build it around the ratio rather than either metric alone.

Limitations, and the dataset

Six properties is a small sample and one of them contributes 89% of the impressions. The blended CTR is really the consumer SaaS site’s CTR with five rounding errors attached, which is why the spread is reported alongside it and why I would treat the median of 0.55% as the more transferable figure.

The verticals skew toward B2B services and software. There is no ecommerce, no publisher, no local business in this set. AI surfaces behave differently on transactional and local queries and nothing here should be read across to them.

A seventh property was excluded. It has Search Console data only from 10 July, less than half the window, so it does not belong in a like-for-like comparison. For what it is worth it recorded 7,981 impressions and zero clicks in that partial period, which points the same way.

One thing worth flagging about the data itself. When this study was first scoped on 27 August the anchor property showed 58,062 impressions for the window. Re-pulled on 2 September, the same window and the same property returned 57,968. Ninety-four impressions were revised away, all of them in the second half. Search Console data settles for longer than most people assume. Re-pull before you publish a number, because I nearly did not.

What would falsify this: a larger sample where question-form queries at good positions do convert; any property where AI impression growth coincides with proportional click growth; or Google shipping clicks in the generative AI report and showing that AI-surface CTR is materially different from the blended rates here. I would genuinely like to see the third one.

The full dataset is published as a single CSV under CC BY 4.0: the six-site table, the two-half comparison, the position-bucket breakdown, and the query-level aggregates behind Findings 2 and 3. It is anonymised at the property level and carries no client-identifiable data. Run the same pulls against your own properties and tell me if you get a different answer.

Frequently asked questions

Does Search Console show AI Mode clicks?

No. Google’s generative AI performance report carries impressions only, broken down by page, country, device and date. Clicks, click-through rate, queries and average position are all absent. The AI clicks do exist inside your ordinary performance report totals, because the generative AI report draws from the same Web search type, but they are not separated out and there is currently no way to isolate them.

What is AI Mode SEO?

AI Mode SEO is optimising for visibility inside Google’s AI Mode and AI Overviews rather than for classical blue-link rankings. The practical difference is that the success metric changes. In classical SEO a top-ten position predicts traffic. On the six properties studied here, 37 queries at an average position of ten or better returned zero clicks, so position stopped predicting traffic and the question becomes whether appearing in an AI answer is worth anything at all.

Are AI Mode impressions worth anything?

On this evidence, very little on their own. Across six properties and 24,498,104 impressions over 92 days, blended CTR was 0.86%, and per-site rates ranged from 0.09% to 0.91%. That is roughly 117 impressions per click at the blended rate, and over 1,100 at the low end. Impressions are a presence signal, not a traffic forecast, and should never be reported as though they were traffic.

Why did my average position get worse when impressions grew?

Because generative AI impressions enter the same Web search type your average position is calculated over, and they frequently appear at deep positions. A flood of low-position impressions pulls the mean down even as visibility rises. On the anchor property here, impressions rose from 4,129 to 53,839 across two consecutive 46-day windows while average position moved from 17.7 to 31.8. That pattern is growth, not a penalty.

How is AI Mode reporting different from AI Overviews reporting?

It is not separated. Google’s generative AI performance report covers AI Overviews and AI Mode together in a single impressions figure and offers no dimension to split them. If you need to distinguish the two surfaces you have to do it outside Search Console, which is what dedicated AI Overviews tracking exists for.