The Generative Blindspot in Google Search Console

Why GSC still can’t tell you which prompts matter — even now that it counts AI traffic

If you work in Generative Engine Optimisation (GEO), you’ve probably noticed something unsettling when analysing Google Search Console.

The question-based, conversational queries that look most like real prompts are largely missing. And when GSC does include them, they arrive stripped of the context that made them prompts in the first place.

This isn’t a content problem. It’s structural — and it has serious implications for how we discover, prioritise and track prompts across the tools we now rely on.

This article covers:

Why prompt-like queries are systematically under-reported, even in GSC’s updated 2026 form
What this tells us about how generative search actually works
Why popular prompt-tracking approaches are struggling
A practical framework for prompt discovery that doesn’t depend on Google’s reporting layer

1. The observation: where did all the questions go?

A common early GEO method has been:

Extract question-format queries from GSC (e.g. “what is…?”, “how does…?”, “which is the best…?”)
Select the top 5–20 by impressions or clicks
Test visibility and answer quality in engines like GPT, Claude, Gemini and Perplexity

This is a logical approach: one would assume as AI search takes off, questions would be plentiful and measurable in GSC.

But practitioners today are seeing:

Very few conversational queries surfacing in GSC
A growing dominance of navigational, branded or truncated keyword strings
An absence of natural-language questions that resemble real prompts

At first glance, this looks like a content or relevance issue. Or, for those of a tinfoil persuasion, a Google plot. It isn’t. It’s an artefact of how GSC reports data — this article has a detailed look.

2. What’s actually changed in GSC (and what hasn’t)

Let’s clear up the picture, because the situation has moved since AI Overviews and AI Mode rolled out.

What GSC now does: Since June 2025, traffic from AI Overviews and AI Mode is included in the Performance report. If your page is cited in an AI Overview, that counts as an impression. If a user clicks through from an AI Overview or from inside AI Mode, that counts as a click. Google updated its official documentation to confirm this.

What GSC still doesn’t do: There is no filter to separate AI traffic from classic organic traffic. AI Overviews and AI Mode are folded into the “Web” search type alongside traditional ten-blue-link impressions. A fake “AI Overviews” filter screenshot circulated in September 2025 and was publicly debunked by John Mueller. A possible “AI contribution pilot” has been referenced in Google support documentation as recently as April 2026, but nothing has launched.

So the original concern — “AI Overview prompts never reach GSC” — was overstated. The current concern is more interesting: GSC sees that something AI-driven happened, but it cannot show you the conversational journey behind it.

The user may have started with “compare best vegan protein powders for endurance athletes”, refined it through three follow-up questions inside AI Mode, and finally clicked through to your page on the fourth turn. GSC logs that final query — possibly anonymised, depending on its uniqueness — and gives you no visibility on the chain of intent that produced it.

That isn’t a small caveat. It’s the core problem. The “query” GSC reports is increasingly a fragment of a multi-turn conversation, not the conversation itself. Again, tin-foil hatters will be twitching that this persistence of last click (hell — any click) benefits Google.

3. The real structural issue: three layers of blindspot

The generative blindspot isn’t one mechanism, it’s three stacked on top of each other.

Layer 1: Anonymisation strips the long tail

Ahrefs’ 2025 analysis of 22 billion clicks across 887,534 properties found that 46.77% of GSC clicks come from queries that don’t appear in the report at all — anonymised because they’re not searched by more than “a few dozen users” over a two-to-three-month window. For smaller or more niche sites, query visibility can drop as low as 0–37%.

This mechanism hits prompt-style queries hardest by design. Conversational queries are longer, more unique, and lower-volume per exact string. They fall below the anonymisation threshold mechanically — not because Google can’t see them, but because privacy rules require it to hide them.

As queries continue to lengthen — and Google has now removed the 32-word cap on search input — the proportion of clicks attributed to anonymised queries is expected to rise further.

Layer 2: The prompt chain is invisible

Even for queries that do surface, GSC reports only the final terminal query — the one that produced the click. The reasoning chain, the refinements, the comparative phrasings the user worked through in AI Mode before landing on your page: all invisible.

The data is there, but stripped of the structure that made it strategically meaningful.

Layer 3: Off-Google prompts never touch GSC at all

And of course the biggest blindspot of the three. Every prompt issued inside ChatGPT, Claude, Perplexity or Gemini — every brand citation those engines deliver — happens entirely outside Google’s measurement surface. Your brand might be cited 12 times a day inside ChatGPT for prompts that have no analogue in Google search at all. GSC will never report it, because Google was never involved.

This also has the potential to massively skew results because already users are turning to different engines for different uses — Claude and Perplexity are often favoured for productivity related tasks, Google AI Mode for commercial queries.

So ultimately, this becomes a condition of the category itself. In a Googlised monopoly, GSC is great for scraping the queries from the Google ecosystem. Where they’re sharing the space with other providers, it’s inadequate.

4. Why this breaks common prompt-tracking approaches

Three knock-on effects worth flagging.

GSC-seeded prompt lists are filtered samples. When you export “top questions” from GSC, you’re not sampling user intent — you’re sampling the subset of user intent that survived Google’s anonymisation threshold and showed up enough times to make the top-1,000 row limit. The selection bias is severe and runs against exactly the queries GEO practitioners care about most.

Volume-based tools struggle by definition. Tools like Semrush and Ahrefs depend on observable query volume. Prompt-style queries often fail both the volume and the observability test. This is why most credible GEO platforms now generate AI-suggested prompts, infer demand semantically, or probe models directly rather than relying on volume estimates. It’s not a flaw in those products — it’s a necessity.

Attribution is fragmenting. Even when GSC does report a useful query, you can’t tell whether that impression came from a classic SERP, an AI Overview citation, or an AI Mode side panel. They all show up as “Web”. So your strongest-performing AI-cited content can look indistinguishable from a stale page ranking position seven for a long-tail term.

5. How to discover real prompts when GSC won’t

If GSC can’t be the primary source, prompt discovery needs to shift upstream. Five practical methods, in roughly increasing order of value.

5.1 SERP feature extraction

Instead of looking at GSC queries, mine the questions Google itself surfaces:

AI Overview prompts and visible follow-ups
People Also Ask expansions
Related questions

APIs like DataForSEO, RankRanger or Serpstat make this tractable at scale. The advantage: these represent live, current intent — not the slow, anonymised, aggregated version that eventually reaches GSC.

5.2 Clickstream and panel data

Third-party clickstream providers (Datos, Similarweb, others) capture user behaviour even when no click occurs and surface question patterns that never reach GSC. Useful as a directional signal, less useful for specific prompt strings.

5.3 Entity- and intent-led clustering

Stop asking “what queries do users type?” and start asking “what decisions are users trying to make?”. Cluster around entities, concerns, comparisons and outcomes — then generate prompt variants from those clusters. This aligns far more closely with how LLMs interpret intent in the first place. At GEO Jetpack our approach is bottom up, not prompt down. What entities is it mission-critical for you to own? Find those and work upwards towards discovering which possible prompts cover those entities, not the other way around.

5.4 Synthetic prompting

Instead of waiting to observe what users searched, use an LLM to generate the prompts users plausibly would issue based on your entity’s knowledge graph and the intent clusters above. This sounds circular but isn’t: you’re using the same kind of model your users are using, with explicit grounding in your domain. The output is a candidate prompt set you can then test.

This is the discovery method that scales. It’s also the only one that produces prompts in volume for emerging or low-traffic topics, where every other method falls flat.

5.5 LLM echo-testing

The validation step. Take the prompts surfaced by the methods above, run them across GPT, Claude, Gemini and Perplexity, and measure:

Whether the model answers at all
Which entities and sources it draws on
Whether your brand surfaces, and how consistently across variants
Which competitors are taking the citation slot you should occupy

If the model responds — and your brand surfaces — the prompt is real enough to matter, regardless of whether Search Console ever sees it.

In practical terms, this is the workflow we run across client GEO campaigns. The pattern that recurs almost every audit: a meaningful fraction of the prompts where a brand can surface in LLM answers have effectively no measurable GSC footprint at all. The brand is winning a visibility battle Google never reported.

6. Prompt generation cheat sheet

Dimension	Classic SEO (GSC era)	GEO (new framework)
Starting question	Which keywords drive traffic?	Which entities must we own?
Unit of value	Clickable URL	Entity citation / brand mention
Primary signal	Query volume, CTR, position	Citation frequency, semantic relevance, model agreement
Success metric	Organic sessions	Share of model, citation share of voice
Discovery method	GSC + keyword tools	SERP feature extraction, entity clustering, synthetic prompting, LLM echo-testing
Measurement surface	Search Console	Cross-engine probes (ChatGPT, Claude, Perplexity, Gemini) + SERP feature extraction
Attribution clarity	Query → URL → click	Prompt chain → model selection → citation → optional click

7. A reframing for GEO success

The key mindset shift is this:

Prompt relevance is no longer proven by search volume or impressions. It’s proven by whether a model can answer the question, which entities and sources it draws on, and how consistently your brand surfaces across variants.

Much of this will never be visible in Search Console. And that’s fine.

GSC isn’t broken. It’s optimised for a world where queries were typed, results were clicked, and URLs were the unit of value. Generative search changes all three. For GEO practitioners, the implication is clear:

Stop treating GSC as a prompt discovery tool. Use it for what it still does well — performance diagnostics, branded demand monitoring, classic SEO health — and build a separate, generative-native methodology for understanding the prompts that actually matter.

Because the most important questions your customers are asking?

They’re increasingly the ones Search Console will never show you in full.