How we measure AI visibility

1. The engines and how we query them

TrendlyAI tracks the three AI surfaces buyers use most when they research a purchase: ChatGPT, Perplexity and Google AI Overviews. We query them programmatically so every scan is asked the same way, on schedule.

Swipe to compare →

For every answer we store the full answer text, whether your brand is named, where it appears, which tracked competitors are named, and every source the engine cited. Those stored answers are what the score, the gap list, the fix drafts and the white-label reports are built from — nothing is estimated.

2. Which questions we ask

A prompt is a question a buyer might type into an AI assistant. When you add a brand, we look up what the domain actually sells (category, description, likely competitors) and generate buyer-intent prompts from that profile. The mix is weighted toward category questions that don't name you, because that is where new customers discover brands:

• “Best [category] for [use case]” — about 30% of prompts.
• “[Category] for [audience]” (small teams, agencies, enterprise…) — about 20%.
• “How to [problem the product solves]” — about 15%.
• Comparison and alternatives questions (“[competitor] vs [brand]”) — about 20%.
• Review and pricing questions (“is [brand] worth it”) — about 15%.

You can edit, remove or add prompts at any time in the dashboard — agencies usually replace a few with the exact phrases their client's sales team hears. The free check uses 5 prompts generated the same way.

3. Three samples per prompt — and why

Ask ChatGPT the same question twice and you will often get two different lists. Answers depend on sampling randomness, on which pages the search step happened to retrieve, and on model updates. Independent research backs this up: a SparkToro and Gumshoe study reported by Search Engine Land found that AI recommendation lists rarely repeat from one run to the next.

So on paid plans and the trial, every prompt is asked three times on every engine in each scan. We report a mention rate — “named in 2 of 3 answers” — instead of a yes/no. In gap lists, a prompt × engine pair counts as covered when your brand appears in at least half of its samples. Three samples is a deliberate trade-off: enough to separate “reliably recommended” from “got lucky once”, while keeping scans affordable at $19–$49 a month.

4. How the 0–100 visibility score is calculated

Every individual answer gets a position weight based on where your brand appears:

Swipe to compare →

For each prompt × engine pair we average the weights of its samples. That number is effectively the mention rate multiplied by how high you were ranked when you were mentioned — named first in 2 of 3 answers gives (1.0 + 1.0 + 0) / 3 = 0.67. The visibility score is the average of those values across every prompt × engine pair, multiplied by 100 and rounded. Per-engine scores use the same formula restricted to one engine.

A score of 100 means you were named first in every sampled answer to every prompt. A score of 0 means no answer named you. Because each prompt × engine pair counts equally, one prompt can't dominate the score just because it was sampled more often.

5. How share of voice is calculated

Share of voice compares you with the competitors you track. Answers are grouped by prompt × engine. In each group, every tracked brand contributes its mention rate (answers naming it ÷ answers in the group). A brand's share of voice is its summed mention rate divided by the summed mention rates of all tracked brands, shown as a percentage.

Your brand counts as mentioned when our answer parser flags it — the same signal used for the score. Competitors count when their name appears in the answer as a whole word (case-insensitive). Only the competitors you list are included, so share of voice is “share among the brands you track”, not share of the whole market.

6. Scan cadence

• Solo: a full scan every week (25 prompts × 3 engines × 3 samples), plus one manual re-scan a week.
• Pro: a full scan every day for each of up to 3 brands (50 prompts per brand), plus one manual re-scan a day.
• Free check / pitch audit: one-off, 5 prompts × ChatGPT + Perplexity, 1 sample each.

After every scan we also check whether pages you published from a fix draft are now cited by an engine for the prompt they targeted. When one is, it shows up as a won citation — the clearest proof for a client that the work moved something.

7. Limitations — read before you quote a number

Answers vary by user, location and model version

Consumer AI apps personalise. A logged-in ChatGPT user with memory, a different country or a newer model may see a different answer than our query did. Engines also update their models without notice, which can shift results overnight for reasons unrelated to anything you changed.

API answers are not identical to the consumer app

We query through APIs and SERP data. The API versions use the same underlying model families and live web search, but the consumer apps can add their own system prompts, shopping modules and personalisation. Treat our results as a consistent baseline that is comparable week to week, not a pixel-perfect copy of any single user's screen.

Detection is text matching

Brand and competitor detection works on names in the answer text. Very generic brand names, or brands commonly abbreviated, can be over- or under-counted. If you see a false match, edit the brand name or competitor list.

Samples reduce noise, they don't remove it

Three samples make a mention rate far more stable than a single answer, but small prompt sets still move. Look at trends across several scans before drawing conclusions — and say so in client reports. Every TrendlyAI PDF report prints its engines, prompt count, samples per prompt and data date for exactly this reason.

Questions

See the method in action on your own domain with the free AI visibility check, compare plans on the pricing page, or read how agencies use it in TrendlyAI for agencies.