How we measure AI search visibility — transparently

One scan = one tracked question, run against every engine your tier covers, parsed by deterministic code, persisted as auditable rows in our database. Four steps:

1. 1 · Send the prompt

   We send the exact question you tracked — verbatim, no extra instructions, low temperature for parsing consistency. Concise system framing (“answer in 2–4 sentences, mention specific products by name”) keeps responses comparable across engines.

2. 2 · Parse the response

   Layer 1 is regex-based: did the response contain the brand hostname or token? Layer 2 is deterministic interpretation — tone, top-list position, citation-by-domain. Long-tail ambiguous cases get an optional gpt-5.4-nano refine pass. Any LLM error falls back to the deterministic verdict, never silently drops.

3. 3 · Extract citations

   We pull every source URL the engine cited and surface them as a 30-day Citations view per question. A separate brand-level extractor names the products the engine actually recommended in prose (“HubSpot”, “Pipedrive”), not just URL hosts.

4. 4 · Persist for replay

   Every scan, every response excerpt, every cited domain is stored. You can re-derive the score from raw rows months later. We track our own LLM spend per scan in the same ledger.

The GEO Score is a 0–100 composite per engine, then a weighted average across engines. The intuition has two layers:

• Mention rate — what fraction of your tracked questions returned an answer that named your brand at all.
• Quality — when you were named, how prominently. We snap quality to four bands {0, 40, 70, 90} mapped to not_mentioned / mentioned / recommended / top_recommended.
• Baseline lift — a 0.40 floor on quality so a single weak mention still gets partial credit; ramping up linearly toward 1.0 as quality improves.
• Engine blend — ChatGPT 0.45 · Google AI Mode 0.30 · Perplexity 0.25. Engines with no results in the scan window are dropped from the denominator.

The exact formula (confidence weighting, edge cases, MVP quality bands) lives in the source — read the full reference in /docs or browse the engineering blog for a deeper walkthrough.

Running every tracked question against every engine every day would burn LLM budget on signal you cannot act on. Two safeguards keep cost and noise honest:

• Cadence floor by tier. Daily pulse for monitored questions, weekly deep cycles for variant runs. Unmonitored questions sit at on-demand cadence until you flip them on.
• Question Confidence buckets. Every tracked question gets a 4-signal trust score (LLM domain coherence + DataForSEO PAA demand + Search Console token overlap + AI scan baseline mention), bucketed as Verified / Worth tracking / Off-target.
• Confidence-weighted scoring. The GEO Score weights each scan result by its source bucket — Verified 1.0, Worth tracking 0.7, Off-target 0.3. A noisy question cannot drag a real signal into the floor.
• Outcome Loop measurement. When you mark an action done, we re-scan the affected questions 14 days later and report the actual GEO Score delta. Lift is measured, not promised.

Most “rank in ChatGPT” vendors lead with a traffic-lift number — “216 % average traffic increase”, “3× faster citations”, “dominate GEO in 30 days”. None of those numbers are falsifiable: there is no public instrument to audit them with. We sell the instrument, so we hold ourselves to a higher bar:

If the answer to “did this actually work?” matters to you, measurement-grounded reporting beats SEO-by-promise every time. That is the entire thesis of this product.