GeoPulse
Methodology

The score, the formula, the caveats.

Most AI visibility tools hand you a number and refuse to show the math. We disagree. Here is exactly how GeoPulse computes the 0–100 score, which engines we poll, what we measure, and where the limits are.

The formula

Score = Mention × Position × Sentiment

Three dimensions, weighted, computed across 20 runs per scan (5 queries × 4 engines), normalised to 0–100.

score = 0.50 × mention_rate + 0.30 × position_score + 0.20 × sentiment_score

All three sub-scores are normalised 0–1, then aggregated and multiplied by 100. Each engine contributes per its weight (see below); per-engine scores are also returned in the report.

50%

Mention rate

Across the 5 queries × 4 engines (20 runs per scan), how often is your brand named at all? A score of 50% means you appeared in 10 of 20 runs.

30%

Position

When you do appear, are you the first option listed, or buried under 3 competitors? We measure ordinal rank within each answer and reward early citations.

20%

Sentiment

Are you described as the recommended choice, a neutral option, or with caveats? We score sentiment per mention against the surrounding clause (positive / neutral / negative).

The 4 engines we poll

ChatGPT · Claude · Perplexity · Google AI Overview

Together these four cover the assistants and AI search surfaces that intercept the majority of buyer-intent informational queries in 2026. Free /check runs ChatGPT + Claude; paid tiers add Perplexity and Google AIO.

ChatGPT (GPT-4 class)

30%

Largest consumer surface for AI assistants. We poll the public conversational endpoint with browsing disabled for reproducibility, then re-poll with browsing for a delta measurement.

Claude (Sonnet class)

25%

Fast-growing assistant share, strong B2B knowledge-worker adoption. We poll without tools for parity with the ChatGPT no-browsing baseline.

Perplexity

25%

AI search engine, heavy reliance on real-time web citations. Treated as the strongest signal for retrieval-based visibility (citations actually link back to your site).

Google AI Overview

20%

Where most informational SERP traffic is being intercepted in 2026. We capture the AIO answer when triggered and record presence + citation slot.

The 5 default queries

Buyer-intent, not vanity prompts

We seed every scan with five templates that mirror how a real prospect actually phrases an AI assistant question. Category & competitor slots are auto-inferred from your domain; paid tiers let you override the list.

  1. 1best [category] for [persona]
  2. 2[your brand] vs [top 2 competitors]
  3. 3[your brand] alternatives
  4. 4how does [your brand] compare with [adjacent solution]
  5. 5is [your brand] worth it for [use case]
Reading the score

What each band actually means

0–24

Invisible

Your brand is essentially absent from AI answers in your category. Competitors are taking the slot. Urgent.

25–49

Inconsistent

You surface on direct-brand queries but lose head-to-head and category queries. Typical for early-stage SaaS.

50–69

Recognised

You appear in most category answers but not first, and sentiment is mixed. Optimisation has clear ROI here.

70–84

Strong

Reliably surfaced, often first, mostly positive sentiment. This is the band paying customers tend to keep tracking monthly.

85–100

Dominant

Category-defining presence. Competitors are described as alternatives to you. Rare — and worth defending.

Principles

Things we will not do

Live API calls, every scan

Every /check run hits the real LLM APIs against your URL. No cached numbers, no demo scores. Run it twice in a row — the numbers will move within normal variance.

Reproducibility over volatility

Temperature is pinned, browsing/tools are disabled for the baseline run, and queries are templatised. Two scans 30 minutes apart should land within ~±5 points absent a real shift.

No optimisation for the score

We do not sell SEO services or guarantee score improvements. The score is a diagnostic. The fixes are yours to ship.

EU data residency

Scan inputs, results and email opt-ins are stored in Supabase Frankfurt. We never train models on your data. DPA available on request.

Limits & caveats

Where the score is wrong

LLM stochasticity. Even with pinned temperature, identical prompts can return slightly different answers. We mitigate by running 20 calls per scan and reporting the aggregate; expect ±3–5 point variance scan to scan absent a real shift.

Browsing bias. Engines with live retrieval (Perplexity, Google AIO, ChatGPT with browsing) reward recent web presence. If your site is new or blocked from crawlers, the score will under-represent your real authority.

Category inference. For free scans we infer your category from your homepage in one pass. If the inference is wrong, the queries are wrong, and the score is meaningless. Paid tiers let you override category + queries + competitors directly.

Sentiment edge cases. Sarcasm, hedged recommendations and conditional mentions are hard to score reliably. We default to neutral when the surrounding clause is ambiguous.

Run the formula on your own brand

Drop your URL. 5 queries × 2 engines on the free tier — your 0–100 score, sub-scores, and the top issues, in about 30 seconds.