Hi everyone, this is Neo.

Let me get straight to the point: Google has officially admitted that the generative AI performance report in Search Console isn’t good enough.

Around September 10, someone on Reddit laid out every flaw in that report, one by one — impression metrics that don’t mean what you think, links behind an expander that go uncounted, position data that answers a question nobody asked, and numbers you can’t add together. Google’s John Mueller replied almost immediately:

“This is pretty much it – we tried to document it as clearly as possible in the help center page… Position for these is hard to do in a way that makes it useful, so we’re currently tracking it like we do for many search features (as a block), & it’s not separated out in the Gen-AI performance report.”

Read that again. That’s Google, at the product level, admitting the data can’t be shaped into what you actually want.

Search Engine Journal picked up the exchange on September 13 and the industry has been chewing on it since. But most coverage stops at “Google admitted it.” Nobody is telling you how to actually read the report now. That’s the gap I want to fill today.

What actually happened

The timeline is short:

• June 3, 2026 — Google announces the generative AI performance report, rolling out to a small subset of UK sites first;
• August 31, 2026 — Google adds one line to its blog post and help docs: “we’ve rolled out these insights to all websites worldwide”;
• Early September — an SEO posts a long breakdown of the report’s problems on Reddit;
• Around September 10 — John Mueller replies publicly, agrees point by point, and adds that “position 1–10 is hard to map”;
• September 13 — SEJ writes it up and the conversation goes mainstream.

The four core problems the Reddit poster raised:

1. An AI Overviews impression doesn’t mean a user ever saw your link;
2. Links hidden behind “Show More” are handled differently, so your real exposure is likely understated;
3. The position metric measures the position of the AI Overviews block, not your link inside it;
4. The Gen-AI report is a filtered view of data you already have, not a new, separate dataset.

Every one of those four changes how you should interpret your own AI visibility. Let’s go through them properly.

Trap 1: An impression is not a pair of eyes

This is the easiest one to trip over, and the easiest to mistake for good news.

Here’s how Google’s own help documentation defines an impression, in substance: an impression is counted whenever your link appears in the page of results that got served — whether or not the item was scrolled into view.

The operative phrase is “whether or not.”

AI Overviews normally sit at the very top of the results page, with several cited links packed into one block. Under the standard rule, if that block renders and your link is in it, the impression is already on your ledger — even if the user never scrolled to it, or clicked a regular result further down.

What does that mean in practice?

• A chunk of your AI impressions are things that nominally appeared but were never actually seen;
• Treating the number as “how many times AI showed me to a person” systematically overstates your reach;
• But it’s not junk data either — it proves your page was selected and rendered by AI. That’s “eligibility to be cited,” not “visibility achieved.”

There’s also a reverse exception buried in the docs: inside independently scrolling or expanding result widgets — carousels, or expandable sections like FAQ results — the item typically has to be scrolled into view, or expanded by a click, before an impression registers.

So within the same report, different result types are counted under different rules. Comparing them side by side without knowing that context is comparing apples to oranges.

Trap 2: Links behind “Show More” get systematically undercounted

The Reddit post put this one better than I can, so I’ll translate it:

The standard rule carves out anything the user has to click to reveal, so links sitting behind that expansion don’t count until someone expands it. Those numbers are understating your exposure rather than inflating it.

Notice the direction — it’s the exact opposite of Trap 1:

• Trap 1 pushes your number up (counts impressions nobody saw);
• Trap 2 pushes your number down (fails to count exposure that was genuinely seen).

One upward bias plus one downward bias, landing on the same metric. Which means: you cannot correct for this with a fudge factor, because neither the direction nor the size of either error is stable.

Practically speaking: if your brand regularly shows up inside the “Show More” expander of an AI Overview, your real AI presence is stronger than the report suggests. Beating yourself up over a modest number in that situation is unfair to your own work.

Trap 3: That “average position” belongs to the AI Overview block

This is the most counterintuitive of the four, and the one most likely to send someone into the wrong meeting with the wrong slide.

An AI Overview occupies a single position on the results page. Every link cited inside it is assigned that same position.

So when you see “average position 1.2”:

• It means the AI Overview blocks your link appeared in sat, on average, at position 1.2 on the results page;
• It does not mean your link was the 1.2nd citation inside that block;
• An AI Overview might list eight citations, and you might be seventh — but if the block is at the top of the page, your “position” is still 1.

Which means the position data in this report only tells you how high on the page the AI block appeared, not where you ranked inside it. And the second thing is the one SEO actually cares about.

Mueller’s reply confirms it: position for these surfaces is “hard to do in a way that makes it useful,” so Google tracks it “as a block.”

One related note on AI Mode: position there follows the same methodology as a regular results page, and a follow-up question inside AI Mode counts as a brand-new query — every impression, position, and click in the new response gets attributed to that new query. We dug into that mechanism back in our August 18 piece on AI conversation queries leaking into Search Console, so I won’t repeat it here.

Trap 4: A filtered view is not extra traffic — never add the two reports together

This one is spelled out most clearly in Google’s documentation and misunderstood most often.

The Gen-AI report is drawn from the existing Web search data in your performance report, filtered to show one slice of it. It is not additional traffic sitting on top of what you already had.

So:

• Never add regular impressions to AI impressions to calculate “total reach”;
• The two overlap by an unknown amount, and you can’t subtract your way out of it;
• The right use is as a share metric: AI impressions ÷ total impressions, tracked over time.

There’s a subtler trap in how the data aggregates, too:

Viewed by property, two results from your site appearing in the same AI feature count as a single impression in the chart total. The chart only splits per-URL once you add a URL filter. And table totals can disagree with chart totals depending on aggregation — that’s expected behavior, not you reading the screen wrong.

Sum all four up and you get the one-line rule: the only trustworthy thing in this report is the relative trend, never the absolute number.

Google admits position 1–10 no longer maps to anything

The most informative part of Mueller’s reply isn’t the admission that the data falls short. It’s the second half:

“Search results pages have a lot of ways for users to interact nowadays, so the old ‘position 1 – 10’ is hard to map, or to make useful for site owners. If any of you have thoughts on what would be useful in terms of tracking position, I’d love to hear & am happy to discuss with the team.”

Google is now openly soliciting the industry’s help on how AI search should be measured. That’s not boilerplate.

Think about what a results page looks like today: an AI Overview block with multiple citations up top, then People Also Ask, a video carousel, shopping cards, a knowledge panel, and then the ten blue links. Now tell me what “position 3” means. The third element? The third plain link?

The concept of position is losing its explanatory power — and that’s not just a Gen-AI report problem. I’ll get into what it means for us in Neo’s take at the end.

Who’s pushing for the missing columns?

Mueller explained why Google can’t build what you want. He said nothing about when. But someone is leaning on the missing part.

First: UK regulators want the click data. In October 2025, the UK’s Competition and Markets Authority designated Google as having strategic market status in general search. Then on June 4, 2026 — the day after these reports launched — the CMA published a final decision imposing a publisher conduct requirement on Google. The CMA’s interpretive notes explicitly call for impressions, clicks, and click-through rates for search generative AI features, reported separately from the rest of general search.

Google has up to nine months to implement all of it, and must report progress every six months during the first year. So far, the separated click and CTR data hasn’t appeared.

One detail that gets lost: clicks are being counted. Google’s documentation states plainly that clicking a link to an external page in an AI Overview counts as a click, and the same applies in AI Mode. The clicks just aren’t broken out in the Gen-AI report. Google isn’t failing to record them — it’s choosing not to show them separately.

Second: Merchant Center got there first. In mid-July 2026, Google opened an “AI performance insights” pilot in Merchant Center, showing retailers which shopping-related questions people ask in AI Mode and AI Overviews. Independent SEO consultant Brodie Clark got access on a client sub-account and published screenshots.

But look at how measured his assessment was — essentially: like the recent Search Console AI rollout, there isn’t a great deal of actionability behind this data, though it’s good to see at least some query-level information, which is exactly what GSC has been missing.

And that query data is grouped, not a raw list of questions. What you get is a vocabulary of your category, not the actual questions anyone typed. Enough to tell you which attributes to add to your product titles and descriptions — but it won’t tell you whether AI Mode sent anyone to your site.

More to the point: the pilot covers a limited set of US accounts, and sites without a product feed — affiliates, review sites, editorial teams — can’t get this data at all. For them, AI search reporting is still the impression-only report in Search Console.

Third, a purely technical snag: this report isn’t in the API.

Neither the Search Analytics API nor the BigQuery bulk export exposes the Gen-AI report data yet. Which means you can’t automate this — you’re exporting CSVs by hand. If you plan to start tracking AI impressions over time, put a manual export on your monthly calendar now and don’t wait for a script.

What independent site owners should actually do (5 moves)

Complaining about the data is pointless. When the data is bad, the methodology has to carry the weight. Here’s what I’d do:

1. Use it for trend reading only — never as a KPI. Build a monthly time series of AI impressions and track its share of total impressions. Keep the absolute number out of your reporting entirely, because it carries both an upward bias (impressions nobody saw) and a downward bias (exposure never counted).

2. Sort by page and pull out your most-cited URLs. Those pages are your AI search storefront. Then check three things on each:

• Any stale data? (AI loves quoting old numbers attached to a year)
• Does the opening paragraph answer the question directly? (The first thing AI grabs is the straight answer up top)
• Are there clear tables, FAQs, and layered headings?

3. Read AI impressions next to branded search volume. Right now this is the only partial workaround for the missing click data. A user can get your brand recommended inside an AI answer, then go search your brand name on Google to buy — and in Search Console that visit shows up as ordinary organic traffic. No tool attributes it back to the AI session.

4. Stop adding the reports together — compute a share instead. Put two numbers in Looker Studio or a spreadsheet: total impressions, AI impressions. Track the AI share by month. That’s the only erosion metric you can compute yourself today.

5. For pages with high AI impressions but flat traffic, check what AI is quoting. If the citation happens to be your most conversion-critical content — a full price table, a spec comparison — then the user read the AI Overview and left without ever needing to click. The fix isn’t “optimize for more impressions.” It’s keeping a hook that requires a visit — deeper test data, a downloadable checklist, a calculator.

Neo’s take

First: the real headline in Mueller’s reply isn’t the admission — it’s the open call for proposals.

“Thoughts welcome” means Google itself hasn’t decided how AI search should be measured. Every third-party AI visibility tool out there — mention rates, cited pages, sentiment — is selling the same thing: the column of data Google won’t hand you, reconstructed with crawlers and prompt sampling.

The awkward part is that in July, Google told CMOs point-blank that third-party AI visibility tools don’t have access to its internal metrics, and named Search Console and Merchant Center as the baseline for tracking gains. So here’s where we are: the official data isn’t good enough, the third-party data has no common yardstick, and the official source is publicly asking for help.

My read: over the next twelve months, the standard for measuring AI search visibility will be forced into existence by regulators and the industry together — not volunteered by Google. Whoever builds a monitoring method they can run consistently, with a fixed definition, comparable month over month, won’t get dragged around when the standard shifts next year.

Second: the collapse of “position” is actually good news for independent sites.

For a decade, ranking anxiety was the core misery of independent site SEO — third today, fifth tomorrow, conversions wobbling with every shuffle. But that whole framework rests on one precondition: the results page is an ordered list.

It isn’t anymore. An AI Overview is a block, and who gets cited inside it, how many links appear, and in what order, is opaque. When Mueller says position 1–10 is hard to map, what he’s really saying is: that ranking curve you’ve been panicking over is losing its explanatory power.

For independent sites, two things need to flip:

One, from chasing rankings to chasing citations. Whether your page gets picked out as a source matters far more than whether it sat at number four. The entry bar for being cited is unglamorous: accurate information, clear structure, verifiable specifics.

Two, from watching position to watching share of structure. With position gone, what matters is how many citations you hold inside one AI answer — and what kind. Are you the primary source or a supplementary one? No tool hands you that today, but you can do it by hand: fix a set of questions, ask the major AIs once a month, and log where you show up and in what role. It’s crude, but it’s more reliable than any dashboard.

Third: this admission was forced, not volunteered.

Look at the pacing. Report launches June 3. The CMA’s final decision lands June 4. Global rollout August 31. And by early September the industry has pulled it apart and Google has conceded it isn’t good enough. That’s not a product roadmap. That’s a schedule set by regulators and public pressure.

Why won’t Google give us click data? My read: if AI search CTRs were published at scale, the “AI is stealing traffic” narrative would be nailed down in public — while Google is fighting copyright and antitrust battles. Click data isn’t just data. It’s evidence. That’s also why the CMA can move them and the industry can’t.

Flip side: as long as that nine-month CMA window is running, you’ll most likely see separated click and CTR data in Search Console sometime in the first half of 2027. When it lands, the monthly baseline you started now is the only thing you’ll be able to compare year over year.

One practical instruction to close on: put “export the Gen-AI report manually, log total impressions” on your operations calendar this month. The report is hard to use today, but it’s the only official paper trail that exists. When the data gets better, everyone else starts from zero — you’ll already have two parallel curves running.

I’ll keep following where this report goes. And as for Mueller’s “I’d love to hear your thoughts” — there’s a real chance some comment on that thread ends up in the next version of the documentation. Worth watching.