Documentation

The problem with asking an LLM once

Anyone can ask ChatGPT "what do you think about company X?" The problem is what you get back: a different answer every time.

Large language models are stochastic by design. Ask the same question twice and the model will shift its emphasis, change its examples, sometimes reverse its opinion entirely. Run a typical GEO/AEO tool twice on the same brand and you'll get two different scores. Upload your pitch deck to a chatbot and ask for a review — you'll get an endless stream of new corrections, forever, because an LLM will never tell you "it's perfect, send it."

That's not measurement. That's noise. And you can't optimize against a target that moves every time you look at it.

Our approach: repetition at scale

There is only one way to extract a stable signal from a stochastic system: ask the same question many times, in many forms, and find what survives.

That's what VeritasLinks does. For every dimension we measure, we send hundreds of structured prompts — the same underlying question, rephrased, reframed, approached from different angles. We count how consistently the models lean one way or another, detect the patterns that repeat across runs, and discard the noise that doesn't. The result is a reproducible score: run the same analysis twice, and you get the same result.

No human analyst could do this manually — it takes over 1,000 prompts per full analysis. That's the point. This is work that has to be automated to be done at all.

Every analysis runs across six leading LLMs — ChatGPT, Claude, Gemini, Perplexity, Grok, and DeepSeek — plus a seventh model (Mistral) acting as an independent validator that cross-checks the findings without participating in the analysis itself.

The pipeline

A full analysis moves through four stages, each feeding the next.

Stage 1 — Digital Footprint

Before we ask any model anything, we establish ground truth. We map the company's actual digital trail: how it positions itself on its own site, what real people say about it across the web, whether reviews exist and what they say, and who its actual competitors are. We filter the noise out of all of this.

This matters because everything that follows is compared against it. Without ground truth, you can't tell whether a model is being accurate, outdated, or simply making things up.

Stage 2 — GEO/AEO Analysis: an interrogation, not a survey

Most tools ask the models a handful of questions and record the answers. We run a full dialogue.

Every model response is compared against the Digital Footprint. When we detect a pattern, a dependency, or a contradiction, we ask follow-up questions — probing the way an interrogator does, watching not just what the model says but how it reasons about the company. Weak sides and strong sides of the brand emerge from this process, and stochastic noise gets filtered out through repetition.

This is also where we separate two things that most tools conflate: being cited is not being recommended. A model can mention your brand and still steer the customer to a competitor. We measure both — visibility and actual recommendation behavior — separately.

Stage 3 — AI Focus Groups

Based on everything learned in the first two stages, we generate five detailed portraits of the company's most likely customers — as the AI itself sees them. Then, for each persona, we test how the models position the brand against its competitors: where it wins, where it loses, and why.

Here's the insight most people miss about AI focus groups: when the AI constructs a persona that is sympathetic to your brand, that persona is a preview of a real person. It's what the actual human will look like who one day asks an AI "recommend me a company that does X" — and gets your name in the answer. You're not reading a hypothetical. You're looking at your future recommended customer.

Stage 4 — Real-time validation

Model knowledge is frozen at training time; markets aren't. Every significant signal from the analysis is cross-checked in real time against live web data through Perplexity — catching discrepancies between what models "remember" about a company and what's actually true today. Your score reflects the current market, not a model's memory.

Pitch deck analysis: the same engine, pointed at your deck

The same AI that builds personas and measures recommendations is the AI sitting between you and the investor you cold-emailed. Most investors now run decks through an AI filter before a human reads them — so we built our deck analyzer to work the way an investor's AI does.

There's a crucial difference between our analysis and uploading a PDF to a chatbot. Given a file, a chatbot reads it as one flat text: it doesn't dig into your claims, doesn't research your market, doesn't check whether your TAM/SAM/SOM numbers hold up — they're just digits to it. An investor doesn't read your deck that way. They know the market, they know the numbers, they sense when something doesn't add up.

Our analyzer evaluates a deck across 12 dimensions, the way an investor's AI would: narrative completeness (problem, solution, why now, traction, the ask), substance quality (specific, quantified, evidenced claims instead of adjectives), traction rigor, market-sizing sanity, business model and unit economics, moat and defensibility, team and founder–market fit, the ask and use of funds, internal consistency (do your numbers match across slides?), communication craft, AI-first-filter readiness (can an LLM extract and summarize your claims without distorting them?), and a dedicated red-flag scan covering the classic deck-killers — inflated TAM, vanity metrics, "we have no competitors," hockey-stick projections with no basis.

Each signal is verified across the same six models and validated against live market data. The deck analysis adds 350–500 prompts on top of the standard pipeline.

What you get

Every analysis produces a report with: your VeritasScore (a reproducible 300–870 score), per-model breakdowns showing how each LLM perceives you, competitor positioning with the reasons behind wins and losses, the specific weak points AI detects, and a prioritized list of what to fix. Every signal in the report includes an explanation of what it means and how to interpret it.

The core promise: a measurement, not an opinion. Run it twice — get the same answer. Fix what it shows you — run it again — see the score move.