Vector Analysis: We Went Where No GEO Tool Has Gone Before
Every SEO tool built for the AI era tracks the same thing: whether your brand gets mentioned in ChatGPT or Perplexity. Mentioned or not. Position in the answer. Sentiment.
That's behavioral data. It tells you what happened. It doesn't tell you why.
We built something different.
The question nobody was asking
When ChatGPT recommends Wise over Airwallex for international transfers, or when Perplexity surfaces Certideal instead of Back Market for a refurbished iPhone query — there's a mathematical reason for that. The model retrieved certain content chunks, compared their semantic vectors against the query, and ranked them by cosine similarity.
That process is measurable. Nobody was measuring it.
Every GEO tool on the market — AIclicks, Profound, Rank Prompt, the rest — runs your brand name through a list of prompts and counts mentions. Useful. But it's reading the scoreboard without understanding the game.
We wanted to understand the game.
What Vector Analysis actually does
Your website isn't just text. The moment a RAG-powered AI looks at your pages, it converts your content into numerical vectors — high-dimensional representations of semantic meaning. Then it compares those vectors against a user's query vector and retrieves the closest matches.
If your top-ranked chunk for "best refurbished iPhone France" is your navigation bar with credit card links — you have a problem. And that problem is invisible to every mention-tracking tool in existence.
Vector Analysis surfaces it.
Here's what the audit covers:
Chunk-level retrieval with cosine scores. We take your actual pages, chunk them the way retrieval pipelines do, embed them, and run your brand's real content against a set of test queries. You see exactly which chunks surface — and what scores they achieve versus competitors on the same queries.
Retrieval experiment across 30 queries. Unbranded commercial queries, split by intent cluster. For each query: whether your brand appeared, in what position, with what sentiment. Across ChatGPT, Claude, Perplexity, Gemini, Grok — depending on your preset.
Competitor chunk benchmarking. Not just "Airwallex ranks above you." We show you that Airwallex's API documentation chunk scores 0.71 on "embedded finance payouts" while your platform page scores 0.43 — and why: because their content has technical depth and yours has CTAs.
DOM noise quantification. Navigation menus, footers, promotional banners, legal disclaimers — these aren't just cosmetic noise. They consume your chunk budget. We measure what percentage of your indexed content is semantically useless for retrieval, down to token counts.
SERM proximity analysis. When your brand appears in an AI answer, what concepts appear nearby? We detect when negative signals — complaint terms, competitor names, hedged language — co-occur with your brand in retrieval results.
Automatic alias detection. TransferWise and Wise are the same brand. If your brand has historical names, transliterations, or product sub-brands, missing them falsifies your retrieval data. We detect aliases automatically from retrieval answers.
Technical surface audit. Sitemap structure, robots.txt crawl permissions, hreflang coverage, structured data gaps, dynamic parameter handling — everything that affects whether AI crawlers can access your content at all.
What we found when we ran it
We've run Vector Analysis across a range of brands during development. A few things became obvious fast.
The navigation bar problem is universal. Marks & Spencer's top-ranked chunk for "best women's cardigans under £50" was their credit card and insurance navigation menu — scoring 0.6358. The actual cardigan content scored lower. This isn't an M&S problem. It's an architecture problem that affects most large e-commerce sites, and no mention-tracking tool would ever catch it.
JavaScript gates destroy retrieval. easyJet's Spanish flight pages return "You need to enable JavaScript to run this app" to crawlers. Nine of eleven pages. Their top chunk for "cheap flights Barcelona to Lisbon" is an unfilled template placeholder: "We're sorry, we don't offer direct flights from to {Destination}." Vueling serves static HTML. The retrieval gap is not subtle.
Brand strength doesn't equal retrieval strength. Back Market dominates the French refurbished electronics market. Trustpilot 4.2/5 from 66,000 reviews. 17 million customers. In our 29 unbranded queries, they appeared in 20 answers — because brand recognition carries weight in model training data. But 89% of their product pages return 403 errors to crawlers. Their competitor Certideal, with a fraction of the brand recognition, wins retrieval on specific product queries because their content is actually accessible.
The VW case flips the expectation. Volkswagen appeared in 27 of 30 unbranded queries — the strongest retrieval performance we've measured. Yet their pages carry 60-70% boilerplate noise and zero JSON-LD schema. They're winning on brand authority embedded in training data, not on retrieval fitness. The moment that advantage erodes — as models retrain on fresher data with better-structured competitors — they'll have no infrastructure to fall back on.
The report format
Vector Analysis comes in three presets.
Fast runs 10 retrieval queries, covers your core pages, delivers the audit within minutes. Designed for a quick read on where you stand.
Balanced scales to 20+ queries, adds competitor comparison, expands the technical surface audit. The right tier for a monthly check-in or a client deliverable.
Deep is the full run: 30 queries across intent clusters, chunk-level cosine scores for every query, full competitor benchmarking, alias detection, SERM proximity analysis, complete technical findings. This is the document you put in front of a CMO or a board.
All three modes support five AI models as the retrieval target: ChatGPT, Claude, Perplexity, Gemini, Grok. Run the same brand against different models and compare where your retrieval gaps are model-specific versus structural.
Reports are generated in PDF, built to be sent directly to clients or internal stakeholders without reformatting.
Who this is for
SEO leads who need to brief developers on what's actually hurting organic AI visibility — not hunches, but chunk scores and token counts.
Marketing directors who need to explain to a CEO why the brand isn't showing up in AI answers despite strong domain authority and a Trustpilot rating in the 4s.
Agency teams who want a deliverable that's meaningfully different from what every other GEO tool produces — because the methodology is different, not just the branding.
CEOs and growth leads who understand that the shift from keyword search to AI retrieval isn't a future event. It's already the traffic pattern for a growing segment of high-intent users, and the brands building retrieval infrastructure now will be harder to displace later.
What other tools won't tell you
Mention tracking tells you that you appeared 7 out of 10 times in branded queries. That's a number.
Vector Analysis tells you that your highest-scoring chunk on "best international money transfer" is a repetitive CTA block, that your mid-market rate transparency content — the thing that actually differentiates you from PayPal — is buried three scroll-depths into a page with 350 tokens of redundant navigation above it, and that Airwallex's API documentation is outranking your platform pages on developer queries because you have "Explore API documentation" links without an actual API documentation page behind them.
That's a diagnosis. And a direction.