Diagnose AI Visibility: The Vector 1 Audit

Reading time: 13 minutes.

Why Diagnostic Comes Before Optimization

The AI search market has produced a strange new pattern. Agencies sell "GEO services" as a single line item, propose a six-month engagement, and start writing content in week one without ever measuring what the engines currently say about the brand they are paid to optimize. The agency is in motion. The diagnostic is skipped. Six months later the client cannot tell whether anything changed because there was no baseline to change from.

That is the pattern Vector 1 exists to refuse. The diagnostic is not a sales artifact, and it is not a checklist tacked onto a proposal. It is a forensic measurement of present state across the four AI surfaces that now mediate buying decisions: ChatGPT, Perplexity, Google AI Overviews, and Gemini. Until that measurement is on the page, optimization is faith.

Two facts shape the urgency. First, AI Overviews now trigger on roughly half of all Google queries, with healthcare and education running far higher and entertainment running lower. Second, and more important for diagnostic strategy, only 38% of AI Overview citations come from pages that rank in the organic top 10, down from 76% as recently as 2024 according to industry tracking by ALM Corp and others. A brand can rank third in classic Google and be invisible to AI search. The two surfaces have decoupled.

Read those figures together and the diagnostic becomes mandatory. If half of Google's traffic is being mediated by an answer engine that draws sources outside the top 10, every brand has visibility it does not see and invisibility it does not see either. Vector 1 is the work that pulls both into the open.

Stage 1: Build the 30-Prompt Test Set

The audit begins with a representative prompt set. Thirty prompts is the working floor. Below thirty, the sample is too small to separate signal from variance. Above one hundred, the data is richer but the labour cost climbs faster than the marginal insight.

The prompts split into three categories of roughly equal weight:

The category prompts are the ones most agencies skip and the ones that produce the most uncomfortable reading for the brand. Brand-direct queries usually return something. Scenario queries reveal whether the brand is competing where the buying decision happens.

Localize where local intent applies. A Brantford mattress retailer needs Brantford, Hamilton, Cambridge, and Kitchener-Waterloo variants of the same scenario prompt because AI Overviews and AI assistants resolve geography differently than classic Google does. A national B2B SaaS does not need that variance, but it does need persona-shifted prompts ("I run a 20-person sales team," "I am an enterprise procurement officer").

Stage 2: Run Across the Four Engines

The four AI surfaces that matter for almost every Vector 1 audit, in order of audit priority:

For each engine, log the prompt, the answer text, every source cited, the position your brand occupies if mentioned, every competitor mentioned, and any factual claim the engine makes about the brand that would need verification. A spreadsheet is sufficient. The labour is the labour. There is no shortcut to running 30 prompts across four engines three times each, which is 360 individual reads, before the dataset is interpretable.

This is also the stage at which the diagnostic refuses the standard agency shortcut of "one screenshot of ChatGPT mentioning the brand, posted to LinkedIn as proof." A single sample tells you almost nothing. The whole point of Stage 2 is the dataset, not the highlight reel.

Stage 3: Score Mention, Citation, Hallucination

Once the dataset is collected, Vector 1 separates three distinct outcomes that almost every shorter audit collapses into one. The distinction matters because each one routes to a different remediation path in the next 11 vectors.

The Aggarwal et al. (2023) GEO paper out of Princeton and Cornell formalized impression metrics for generative engines, including Word Count and Subjective Impression measures that capture how prominently a citation appears inside the answer rather than whether it merely exists. A citation buried as the eighth source has different downstream value than a citation pulled into the visible answer text. A serious audit captures this gradation, not just the binary cite-or-not.

Hallucinations are the asymmetric finding. One bad fact repeated across thousands of prompt instances per month does measurable revenue damage, and unlike rank, hallucinations do not gradually decay; they compound until the underlying entity record is corrected. Vector 2 (Anchor) is where most hallucination remediation happens, but Vector 1 is where the targets are surfaced.

Stage 4: Account for Non-Determinism

The single technical fact that separates a credible Vector 1 audit from a marketing-deck audit is non-determinism. Generative engines do not return the same answer twice. A prompt run at 09:00 may surface your brand and cite three Tier-1 sources; the same prompt run at 09:05 may surface a competitor instead. Aggarwal and colleagues documented this variance directly: even with identical prompts, identical models, and identical session state, output diverges. Industry observation suggests the chance of identical responses on consecutive runs of the same prompt is below one in one hundred.

This is not a bug to engineer around. It is a measurement reality that has to shape audit design. Three implications follow:

First, every prompt runs a minimum of three times across separate sessions. A prompt that surfaces the brand once and not the next two times has a 33% mention rate, not "yes, mentioned." Second, dates of measurement matter as much as the measurements themselves; the audit is a snapshot of model behaviour on a specific day and must be redated to be defensible. Third, the metric the audit reports is a probability distribution, not a binary. Saying "the brand appears on 70% of category-prompt runs in March 2026" is honest. Saying "the brand appears on Perplexity" is meaningless.

This is also why a Vector 1 audit cannot be replaced by a one-off tool screenshot or a single ChatGPT session. The dataset has to be statistical or it is anecdotal.

Stage 5: Translate Findings Into a Vector Map

The deliverable that makes Vector 1 useful instead of merely interesting is the Vector Map. Each finding from Stages 3 and 4 routes to one of the next 11 vectors, with a priority and a sequencing dependency. The map is what turns the diagnostic into a work plan.

The map is what most audit reports lack and what makes Vector 1 the operational starting point of the methodology rather than a standalone deliverable. A finding without a remediation route is just an observation. A finding with a vector tag is a unit of work.

If your domain has never been formally diagnosed across the four AI surfaces, that is the work that has to land first. A no-charge Vector 1 baseline takes about a week and produces the map the next 11 vectors run against.

What Comes After Vector 1

The 12 Vectors are sequential by design. Each one inherits work from the prior vector and produces work for the next. Vector 1 hands off to Vector 2: Anchor, which is where the hallucination findings get fixed at the entity layer through Knowledge Graph submissions, schema correction, and NAP consistency work. Vectors 3 through 6 then build the resonance, embedding, citation, and structural foundation that turns a corrected entity into a discovered one. Vectors 7 through 10 do the distribution, freshness, clustering, and local work. Vectors 11 and 12 close the loop with measurement and iteration.

The reason the order is fixed is that each vector consumes the output of the prior. You cannot anchor an entity you have not diagnosed. You cannot embed answers for prompts you have not mapped. You cannot measure improvement against a baseline you have not captured. The methodology is a system, and Vector 1 is the system's intake stage.

Frequently Asked Questions