How we measure

Fixed questions, three runs, five AIs, and a count of how often you are named. Why one screenshot proves nothing, and where the method has limits.

By Dasha. Updated November 1 2026. https://heardofyou.io/how-we-measure

We ask five AIs the same fixed questions, three times each, and count how often you are named. Then we do it again, the same way, at day 30, 60 and 90. The before and after is a fair comparison, not a lucky screenshot.

First we check. Then we work. Then we check again.

Why one screenshot proves nothing

Ask an AI "what is the best invoicing app for a freelancer?" three times in a row. You will often get three different lists.

In a study of 2,961 prompts across ChatGPT, Claude and Google AI, the same brand list came back twice less than 1 time in 100. The order, the length and the names all moved.

So a screenshot of ChatGPT naming you is nice. It also tells you almost nothing. So does a screenshot of it ignoring you. Only a pattern across many runs means something.

That is also why we never report a rank. "You are number 2 in ChatGPT" describes one answer that will not happen again.

What we count: mention rate, not rank

For every answer we log, we ask one question: did the AI name you? Mention rate is the answers that name you, divided by all the answers we logged.

Named in 12 of 60 is a mention rate of 20%. We report the count and the rate, never a score out of 100 and never a position.

How we run it

1. Fixed questions, written once. For the free check we write 5 buyer questions in your buyers' words. They cover the category, the use case, a price limit, the competitor you lose to, and a beginner's question. The audit and project use 25. You can edit them before we run. After that they are locked, because a new question gets a new answer.
2. Three runs per AI, in fresh chats. Each question goes to ChatGPT, Gemini, Claude, Perplexity and Muse (Meta's AI app) three times. Every run is a new chat, memory off or logged out, from the same location. No "remember I like Ledgerly" leaking between runs.
3. The same thing again at day 30, 60 and 90. Same wording, same AIs, same number of runs, same location. Side by side, with every answer saved.

Five questions x 3 runs x 5 AIs is 75 answers for the free check. Twenty-five questions x 3 runs x 5 AIs is 375 answers for the audit and for each project re-check.

Which AIs, and why these four
- ChatGPT and Gemini. Consumers use these two most for purchase research. In one survey of 550 AI shoppers, 60% used ChatGPT and 63.5% used Gemini.
- Claude. People who build software, and buy tools for building it, use it heavily. Among 2,000 startup founders, 59% paid for Claude, more than any other assistant.
- Perplexity. It searches the live web for every answer. So it is the first to show a new source mentioning you.

Do your buyers ask somewhere else, like a shopping assistant inside a marketplace? Then we tell you honestly: this check is not the right tool yet.

What we record for every answer

| Field | What it means |
|---|---|
| Named | Yes or no. Did the answer mention you by name? |
| Position | If named, were you first, second, third or later in the list. Logged for context, never reported as a rank. |
| Tone | Recommended, mentioned neutrally, mentioned with a caveat, or wrong. "Wrong" means the AI said something untrue about you. That is its own kind of problem. |
| Competitors named | Every other product in the answer, so we can show who AI names instead. |
| Sources cited | The pages and sites the AI linked or leaned on. These become the target list. |
| The full answer | Saved as text, so you can read it yourself. |

Cited is not the same as recommended

An AI can quote your page and still recommend a competitor. "Ledgerly's blog has a good explainer on late fees. For actual invoicing, most freelancers use FreshBooks."

That answer cites you. It does not recommend you. We log it as a mention with a caveat. We report "recommended" separately, because that is the one that brings customers.

Where the AIs read, and why that matters

Most of what an AI cites comes from third parties, not from the brand itself. In one analysis of 207 million citations, more than 84% came from earned or third-party sources. About 0.3% came from paid content.

Being found in the AI's own live web search looks like the single biggest factor. Brands missing from those results were mentioned about 3% of the time. Brands whose site was found, and searched for by name, were mentioned over 90% of the time.

Ordinary search ranking still counts too. 58% of pages at position one got cited, against 14% at position ten.

So most of our work happens off your site: on the review sites, lists and threads the AIs read. And we record the sources in every answer.

Honest limits
- Small samples are noisy. Sixty answers is enough to see a pattern, not to split hairs. If two checks differ by a couple of answers, we say "no clear change". We do not claim a win.
- Apps and APIs can differ. Where we can, we run checks the way a buyer would: in the consumer apps. API answers can differ a little from the app. We always say which we used.
- Location changes answers. ChatGPT follows the location your device reports. Gemini follows the network's country. We fix one location and keep it for every check.
- Some figures here come from vendors or pre-prints. If a study was run by a company that sells AI visibility tools, or is not yet peer reviewed, the source list says so.
- Mention rate is not revenue. Being named more often is the goal we can measure. If you share your analytics, we also report AI referrals. We never claim revenue we cannot see.
- Answers move for reasons that have nothing to do with you. Models get updated. Sources gain or lose citation share. A rise or a fall can be partly weather. Another reason for three checkpoints, not one.

Everything is saved

Every answer from every run is saved as text and shared with you. Disagree with how we coded one? Read the original. We would rather you check our work than take our word for it.

Sources
- SparkToro, AIs are highly inconsistent when recommending brands, 2026 (https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility/)
- Product.ai survey of 550 AI shoppers, 2026 (https://finance.yahoo.com/technology/ai/articles/product-ai-survey-finds-four-120000400.html)
- Supabase, State of Startups 2026 (https://supabase.com/state-of-startups)
- Muck Rack, Generative Pulse: earned media drives more than 84% of AI citations, October 2026 (https://www.globenewswire.com/news-release/2026/10/28/3390134/0/en/generative-pulse-earned-media-still-drives-more-than-84-of-ai-citations.html)
- Tannenbaum, a fitted stage model of brand mentions in AI search, 2026 pre-print (https://arxiv.org/abs/2609.23162)
- AirOps and Kevin Indig, is query fan-out relevant, 2026 (https://www.getpassionfruit.com/research/is-query-fan-out-a-marketing-buzzword-or-something-relevant)