Insight Theater: Inside Google's AI Prompt Leak
TL;DR: Earlier this month, Google Merchant Center briefly displayed the raw prompt behind its "AI performance insights" instead of the insight itself. Read like an engineer, that prompt is a confession: the model isn't analyzing your data — code already did the analysis. The model's entire job is to write one carefully constrained sentence, with the numbers stripped out, so the dashboard feels like it talked to you — the same trick inside most SEO and GEO tooling sold as "AI insights." If your vendor's AI summary can't survive its own prompt being made public, it was never an insight. Here's the test that separates analytics from theater.
James here, CEO of Mercury Technology Solutions.
Hong Kong — September 2026
What Google's Leak Actually Exposed
Nothing sensitive. That's what makes it beautiful.
A merchant named Vipul Kumar opened Merchant Center's performance dashboard and, thanks to what looked like a rendering glitch, saw the machinery behind the "Top products that drove your performance" card. Not the polished sentence — the instructions for writing it. Barry Schwartz covered it at Search Engine Roundtable on September 7, and this week Apoorv Sharma surfaced it on Threads with the framing nobody else got right: every company using AI-generated dashboards needs to ask one very uncomfortable question.
Here's what Google's own dashboard was instructed to produce. A performance summary where:
The role is "helpful and professional business analyst"
The data is already computed: total product clicks up 1,145, plus a ranked list of 20 products with their click changes
The response must be exactly one sentence
It must use one of three provided templates
It must mention "in the last 28 days"
It must reference the click change without using any numbers
It must highlight the product with the largest change in absolute value
Product titles longer than 10 words get truncated
And it must never, ever use the word "which"
Look at that list again. The ranking was done. The math was done. The anomaly detection was done. Plain, boring, deterministic code had already computed everything a merchant needs to know. The LLM's entire job was to convert "+1,145 clicks, driven primarily by Product X" into one numberless, which-free, template-conforming sentence.
That's not an analyst. That's a copywriter with a compliance department.
How Every AI Dashboard Actually Works
Most coverage treated this as a curiosity — ha, Google showed its homework, funny glitch, moving on. Wrong reaction. This is the most honest analytics UI ever shipped, and it told you exactly how "AI insights" get made across the entire industry.
The architecture is staring at you:
Computation layer — code. Deterministic, testable, auditable. Does the actual math.
Selection layer — code. Picks what matters: top 20 products, biggest absolute change.
Narration layer — an LLM. Takes finished analysis and renders it as prose, under constraints strict enough to make a legal team blush.
並非是分析，而是敘事 — it is not analysis; it is narration. The Chinese compresses it perfectly, because the distinction matters more than the English lets on. And the narration layer is the only part anyone reads.
The Chinese have a phrase for what this layer adds: 錦上添花 — flowers embroidered onto finished brocade. The brocade (your data, your metrics, your trends) was already complete. The AI stitches a decorative flower on top and the vendor bills you for "AI-powered insights."
Here's the detail that should genuinely bother you: the summary must describe a click change without containing numbers. Read that constraint until it sinks in. The single most information-dense artifact in your analytics stack — the delta, "+1,145" — was deliberately removed before the sentence reached your eyes. Compressed from signal into vibe. That's not a bug in Google's prompt. That's the product design.
The Vanishing Test: If every AI-generated summary on your dashboards vanished overnight, name one decision that would change. If you can't, you don't have analytics. You have Insight Theater.
Orientation, Outsourced
John Boyd built the OODA loop around a brutal observation: the side that cycles Observe → Orient → Decide → Act faster wins, and the loop breaks at Orient, not Observe. Everyone is currently buying Observe — more dashboards, more instrumentation, more feeds. Google's leak shows you what happens to Orient when nobody's looking: it gets handed to a one-sentence template with the numbers stripped out.
You didn't automate orientation. You outsourced it to a formatting spec you were never allowed to read.
Yang Wen-li spent an entire war winning because he read the raw reports while the Alliance's high command consumed digests of digests — pre-shaped summaries written by people who had already decided what the data meant. The fleet-level decisions were made on sentences, not numbers. The sentences were wrong often enough to lose a war. Every organization running "AI summaries" without knowing the prompt is the Alliance high command, congratulating itself on how efficiently it stays informed.
The 3-Question Test for Every AI Dashboard Vendor
This is where the uncomfortable question becomes a procurement weapon. When a vendor tells you their platform has "AI-generated insights," stop nodding. Ask three things:
What does the model compute that your code didn't? If the answer is "it summarizes what we already computed," you're buying the narration layer. Fine — narration has value — but price it as UI, not as intelligence.
Where did the numbers go? If the AI summary omits or smooths the actual figures — as Google's deliberately does — ask why a feature called "insight" is engineered to reduce information density.
Can I see the prompt? This is the new "show me your methodology." A vendor whose insight collapses when the instructions behind it become public was selling decoration. If the summary falls apart when you read the prompt, it was always bullshit — you just hadn't seen the ingredients.
Stop asking what the AI can summarize. Start asking what decision the summary changes.
None of this means AI doesn't belong in dashboards. It means the trust architecture has to be explicit. The computation layer can stay closed — it's deterministic, testable, boring. The narration layer shapes what your team believes, which products they push, which fires they fight on Monday. Prompts that shape decisions are product, and product gets audited. This is the same principle we apply to GAIO work with clients: the answer layer — the thing that speaks to humans on behalf of your data — is where trust lives or dies, and it doesn't get to hide behind a magic trick.
Answer engines don't cite vibes. They cite numbers. Strip the figures from your insight layer and you've stripped your citability along with them.
This isn't academic for us. Mercury's founding number: 43% of enterprise leads fall through the gap between what the machine says and what the human actually needs. That gap starts exactly here — in dashboards that mention growth without the number, and in teams that trust the sentence more than the data.
The New Reality for AI, SEO, and GEO
Google will patch the glitch, the screenshot will fade from memory, and a thousand vendors will keep shipping one-sentence miracles with the numbers amputated. You now know exactly what's behind that curtain, because Google — for one accidental moment — showed everyone.
This is what digital transformation actually looks like in 2026 — not new dashboards, new trust architecture.
The companies that win the next five years of analytics, SEO, and GEO won't be the ones with the most dashboards. They'll be the ones who can answer two questions without flinching: what did the machine decide to show me, and what did it decide to leave out? The prompt is the answer to both. Demand to see it.
The brocade is yours. Don't let anyone sell you the flowers without showing you the thread.
Next Blog Post: "The Trust Layer: Why Prompt Transparency Becomes a Procurement Requirement in 2027"
Mercury Technology Solutions: Accelerate Digitality.