Google admits Search Console limit: 5 workarounds for AI Search

Google’s admission is out

I’ll tell you something agencies hate to hear. Google just admitted that its Search Console AI report does not measure what most SEOs think it measures. Not a rumor. Not a marginal blog post. A response from John Mueller, reported by Search Engine Journal on September 13, 2026.

The context? Since August 31, 2026, the Gen-AI report is accessible to everyone. It shows impressions, clicks, positions for appearances in AI Overviews and AI Mode. On paper, we manage AI Search with data. In practice, the dashboard poorly reflects a site’s real visibility.

A Reddit user broke down the mechanism. Google responded. The exact quote from John Mueller, translated from Search Engine Journal: « That’s about right. We tried to document this as clearly as possible in the help center. The position for these results is difficult to make useful, so we track them as a block. »

An admission. Finally.

The Gen-AI report is a counter of appearances in a block. Not a positioning report. The data is already included in the Web report, not in addition. If you add your Web impressions and your AI impressions, you’re counting the same thing twice. That’s the first illusion.

Who manages an AI Search strategy on an impression counter that doesn’t measure real visibility? Nobody serious. Yet, I look at 15 sites per week. Many have already connected a dashboard to this report. They make budget decisions on numbers that underestimate or overestimate their exposure.

Understood? No.

What Google admits is that position doesn’t have the same meaning as for blue links. The AI Overview block is at position 2? All links cited inside inherit that position 2. You think you’re second. You might be the fifth citation in a panel nobody expands. Or the opposite.

The problem is the data architecture.

The report tells you how many times a page appeared in an AI panel. It doesn’t tell you if the link was seen. Or if it was clicked. Or where it was actually placed among the other citations. Yet, in SEO, the position of a link within a set changes everything. The second citation gets more clicks than the fourth. The report doesn’t show that.

That’s why brands must stop basing their AI ROI on incomplete dashboards. The admission comes from Google. The system’s own creator says the indicator is insufficient. What do we do from there? We build other signals.

Concrete signals:

• Server logs for real clicks
• Citations observed on control queries
• Self-declared conversions
• Reporting by semantic block

These five workarounds operate independently of the Gen-AI report. They replace it.

The three illusions of the Gen-AI report

Managing without reliable data starts by naming the illusions. Search Console’s Gen-AI report has three. Let’s go through them one by one.

First illusion: an impression does not mean the user saw the link. Google’s standard rule counts an impression as soon as the element appears on the served results page, whether visible on screen or not. If your link is in a rendered AI Overview but the user doesn’t scroll to it, the impression is counted. You think you’re seen. You’re just present in the code.

Second illusion: links hidden behind « Show more » are underestimated. Google’s rule excludes anything the user must click to reveal. As long as nobody expands the panel, the additional links don’t count. Yet, these links may receive more real exposure than the report shows. The number is sometimes below reality. Sometimes above. Impossible to know which.

Third illusion: the attributed position is that of the block, not the link. Each link inside an AI Overview receives the position of the whole. If the AI Overview block is displayed in second position on the page, all links cited inside have an average position of 2. The report does not distinguish the first cited link from the fifth. You don’t know if you are the main source or a supplement at the bottom of the citation.

Search Engine Journal reports that John Mueller confirmed it in black and white. The position is difficult to make useful, so Google tracks results « as a block ». The measurement model doesn’t hold, and Google acknowledges it.

You don’t replace ten blue links with an AI panel while keeping the same position logic. Ten blue links are hierarchical. The AI panel is an aggregate. Applying the same metric to both is like comparing a categorized library and a discussion thread. None of that helps with decisions.

Here’s the trap: comforting data, and false.

The most annoying part, at its core, is the reading we make of it. A marketing director sees 12,400 AI impressions in the month. He concludes that AI works. He increases the budget. Meanwhile, server logs show 43 real clicks. Order of magnitude observed with my clients.

The useful question is not how many impressions, but how many times your domain was cited, seen, then clicked. The report only answers the first.

Precise. Cold. Useful nonetheless.

These three illusions stem from a single design choice: reusing the ten-blue-link model for an object that has nothing to do with it. Blue links are an ordered list. The AI panel is a generated conversation. Wanting to assign a unique position to a citation within a conversation is like asking the place of a word in a paragraph. The answer exists, but it means nothing.

Concrete result: entire teams optimize metrics that don’t match real behavior. We cut pages that generate clicks because their impressions are low. We invest in pages that impress but don’t convert. The dashboard ends up being a problem.

To escape this trap, one simple rule: never make budget decisions on the Gen-AI report alone. Never. Always cross-check with an independent source of truth.

Voici les cinq parades que je déroule, dans l’ordre où elles s’enchaînent. Chacune répond à une illusion du rapport Gen-AI, de la mesure brute jusqu’au reporting utile.

Workaround 1: server logs, your only clickable truth

An e-commerce client calls in late September. 800 products, a clean catalog, organic traffic growing slowly. He shares his Gen-AI report: 1,240 AI impressions over the last 30 days. He’s satisfied. He wants to double content production.

We ask him to extract raw server logs. Not Google Analytics. Not Search Console. The logs.

Result: 17 real clicks from AI surfaces. Order of magnitude. Seventeen. Not 1,240.

The difference doesn’t come from a bug. It comes from the definition of impression. The server log records a real HTTP request, triggered by a browser. The Search Console impression records an inclusion in a served page. It’s not the same thing.

First workaround: connect your logs to an analytics tool. If you don’t have access to logs, ask your host. Cloudflare, Apache, Nginx: all can provide raw files. You filter bots, you keep human requests, you cross-reference with Google referrers.

How to isolate clicks from AI? Several possible signals: the q parameter in the referrer URL, the user-agent field, the landing page. Clicks from AI Overviews often go through the classic Google domain, but with specific query strings. You have to dig. It’s tedious. It’s reliable.

For this client, once the logs were extracted, we found that 9 of the 17 clicks came from a single page. A comparison page, not a blog article. The Search Console report highlighted 34 pages with impressions. The logs said: 1 page generates 53% of AI clicks.

We stopped producing superficial content. We consolidated the comparison page. We reinforced its semantic architecture.

Three months later, the logs showed 142 AI clicks per month. Still not 1,240 impressions, but real traffic had multiplied by 8.3. Without advertising.

The server log is the physical trace of what happened. The Search Console report says what Google chooses to display. The log says what the visitor did.

Why do so few people use it? It’s technical. It’s less sexy than a dashboard. You have to cross-reference large files. But it’s the only workaround that measures the click, not the impression.

Concrete steps:

• Extract 90 days of raw logs
• Filter bot user-agents with an up-to-date list
• Spot referrers containing /search or /url
• Map landing pages to queries using parameters

You have logs. Why aren’t you reading them?

Workaround 2: observable citations, not impressions

The Gen-AI report measures appearances. What I manage is citations. A citation is the moment an AI panel mentions your domain, your brand, or one of your pages. It’s observable. It’s workable.

The method fits in one line: define 47 control queries. Not 10. Not 100. 47 precise queries, spread across your three priority semantic clusters. Every Monday, query Google, AI Overviews, and AI Mode with these queries. Note if your domain is cited, at what place in the block, with what text.

That’s little data. But it’s stable. And stability is what reveals evolutions. If your domain moves from citation 4 to citation 2 on a high-volume query, that’s a signal. The Search Console report won’t show it. Your manual tracking will.

What tools? A spreadsheet is enough to start. Then, you can automate with visibility tracking solutions that accept AI surfaces. Some offer daily exports, others don’t. The tool matters little. What matters is the regularity of the tracking.

For a natural health client, I set up this protocol on 47 queries. After 6 weeks, 12 queries triggered a citation of his domain. The Search Console report showed impressions on 38 queries. 12 real citations versus 38 theoretical appearances. The gap jumps out.

Citation tracking also allows spotting citations without clicks. That’s useful. A non-clicked citation means your page is visible but not attractive enough. You change the title, the summary, the meta description tag. You watch the effect on the click-through rate of subsequent sessions.

Workaround 2 is therefore to create your own AI visibility dataset. Not a Google dashboard. A manual tracking, at a fixed frequency, on chosen queries. It’s field work. But it’s the only way to know if your semantic is cited, not just displayed.

Which queries to choose? Those that already generate organic conversions. Those that are informational and prepare the decision. Those of comparison. Not volume keywords. Value keywords.

You want to manage AI Search? Stop looking at impressions. Count citations.

Un cocon sémantique, c’est un jeu de pages reliées par du maillage interne autour d’un cluster de requêtes. Voici comment la structure s’organise concrètement.

Workaround 3: semantic cocoon as a positioning compass

When Google says position isn’t measurable, it buries the concept of position. What counts now: the probability that an AI panel takes your page as a source. This probability depends on topical authority. And topical authority depends on semantic architecture.

A semantic cocoon is a set of pages linked by internal links, around a cluster of queries. Thoughtful interlinking, not random. An architecture that tells Google which pages are references on each theme.

Since 2016, I’ve delivered 1,300 semantic cocoons to over 650 clients. I always see the same thing: a site with a clear semantic architecture collects more citations in AI surfaces than a site with scattered content, even when the latter publishes more.

Why? The model that powers AI Overviews doesn’t look for the page best optimized on a keyword. It looks for the source that best covers the semantic field of the question. A cocoon organizes this coverage. It gives the engine a map of the territory. And this map weighs heavier than the rest.

A natural health client had 214 blog articles. No hierarchy. Articles competed on the same queries. AI panels sometimes cited his domain, without regularity. We restructured the content into 8 thematic cocoons. We created pillar pages, support pages, contextual links.

Result after 5 months: 19 citations observed per month at the start, 155 at the finish. The Search Console report showed +820% AI impressions over the same period. Both numbers are true. Only one measures the visibility actually gained.

The cocoon does more than inflate citations: it makes reporting possible. Group your pages into clusters, and you manage AI performance cluster by cluster. You see which theme is cited, which is ignored, which is progressing. Without a cocoon, you manage isolated URLs. With one, territories.

An effective cocoon rests on three pillars:

• Pillar pages that cover broad intent
• Support pages that answer precise questions
• Contextual links that connect the two levels

Workaround 3: make architecture your positioning compass. The Gen-AI report doesn’t give you a reliable position? Build a structure that makes position secondary. Google cites sources, not places. Become a source.

It’s the difference between being listed and being summoned.

Warning: a semantic cocoon is not reduced to cross-links. It’s a mapping of intents. Each page has a precise role: the pillar page captures broad intent, support pages capture specific questions, contextual links transfer authority. If this hierarchy is absent, the cocoon doesn’t take off.

In AI surfaces, this hierarchy weighs even more. Language models use the site’s internal coherence to choose their sources. A site that covers a topic exhaustively and in an organized way has more chances of being cited than a site with a single long article, even a very good one. Coverage counts as much as quality.

The cocoon transforms your site into a structured knowledge base. That’s what AI panels are looking for.

Workaround 4: self-declared conversions, the only ROI that counts

Impressions don’t pay your bills. Neither do clicks. What pays are conversions. To manage AI Search without reliable position data, connect your measurement to conversion, not to appearance.

The simplest workaround: ask your visitors how they found you. A field in the contact form. A dropdown with the options: « Classic Google search », « Google AI answer », « ChatGPT », « Recommendation », « Other ». It’s self-declared. It’s imperfect. But it’s a source of truth Search Console will never provide.

For a B2B client, this field was added in 20 minutes. In 8 weeks, 37 leads checked « Google AI answer ». The Gen-AI report showed impressions on 61 pages. The conversions said: two pages generated 31 of the 37 leads. I concentrated efforts on those two pages. I reduced useless production elsewhere.

Why does it work? The visitor’s memory is more reliable than the referrer. The referrer can be truncated, hidden, absent. The visitor remembers the channel, especially if the AI answer was decisive. Accuracy is not perfect, but the trend is usable.

Add UTM parameters on the links you control: newsletters, partnerships, Google Business Profile listings, awareness campaigns. Clicks from AI Overviews don’t always carry a UTM. But some call tracking and form tools capture the declared source. Enable call transcription. Listen to the first minutes. Clients say it: « Google’s AI recommended your comparison page ».

Self-declared conversion has an advantage logs don’t have: it measures realized intent. A click doesn’t tell if the visitor found what they were looking for. A qualified lead does.

Build a dashboard with three columns: declared source, landing page, lead value. After 90 days, you’ll know which AI surfaces generate revenue. Not impressions. Not clicks. Contracts.

The Search Console report will never tell you how many signed quotes come from an AI Overview. Your forms will.

Workaround 4 is the shift in reporting from visibility to value.

If you sell high-average-basket services, also connect call tracking with dynamic numbers. Each source receives a different number. The call is recorded, timestamped, linked to the form. You know precisely which AI panel generated the call. It’s heavier to install. But when a contract is worth $25,000, the precision pays for the tool in a week.

One last point: train your sales teams to ask « how did you discover our company? » at the beginning of the exchange. The spontaneous answer is often richer than the dropdown. It gives the exact query, the tone of the AI answer, and sometimes the precise citation. This verbatim becomes first-rate qualitative data to adjust your content.

Workaround 5: reporting by block, not by URL

Google says it: position is tracked by block. Following URLs one by one wastes your time. Follow blocks of pages. Clusters. Themes.

A client had 945 blog pages. His Search Console dashboard listed 945 rows. Unreadable. We grouped pages by search intent and semantic interlinking. Twelve blocks. No more.

Block reporting starts from an observation: in an AI panel, what is cited is the site’s ability to answer, not an isolated page. A single page is never cited stably. It’s the block of linked pages that covers the field and attracts the citation.

The method: export your Search Console data via the API. A regex on the page paths to map them to your 12 blocks. Aggregate impressions and clicks by block. Add log data by block. Add self-declared conversions by block. You get a management dashboard by semantic territory.

For this client, the « hardware comparison » block represented 72% of AI clicks measured in logs, but only 22% of Search Console impressions. The « beginner guides » block showed massive impressions but almost no clicks. By stopping superficial guides and reinforcing the comparison block, real AI traffic multiplied by 4.6 in one quarter. Order of magnitude.

Why does it work? Because your reporting aligns with how Google treats AI: through aggregation. Instead of fighting the data limit, you leverage it.

The Gen-AI report gives you impressions per page? Transform them into signals per cluster. The page is a sensor. The cluster, the indicator. We move from a granular view to an architectural view.

Three steps:

• Map each URL to a semantic block via a documented regex
• Aggregate metrics by block, never by URL alone
• Compare blocks among themselves to allocate content budget

Block reporting changes the conversation. We no longer ask « is this page positioned? ». We ask: « is this semantic territory gaining citations, clicks, conversions? ». That one, we can answer, even with imperfect data.

And you, your next Search Console report, what will you interpret it with: faith or logs?

Stéphane Jambu

SEO & AI Engineer

I build growth systems / AI / Neuroscience | 650+ clients · 80 LinkedIn testimonials · 30 years of expertise · 15 years of systems running without me.

Follow on LinkedIn