Conversion intent

A journey tells you where people got stuck. Conversion intent adds one more question at the end: having been through it, would they take the action you care about? It is an optional read of how a simulated audience leaned toward one event, buying, signing up, or subscribing, and it is a direction to compare versions by, not a forecast.

​

Each person answers in words, not a number

A study opts in by naming the event in evaluation_dimensions. Each value adds one open question that every participant answers after the run, alongside any other after-run questions: The simulated person answers in their own voice, with the whole session behind them. Nobody picks a point on a scale, so the read keeps the reasons and the hedges behind each answer. The question is kept in step with the setting: turn the event off and it is removed, turn it on and it is added, and your own questions keep their order. It does not appear among the questionnaire answers in the web app or in exports, because its answers feed this read instead.

​

From words to a five-point spread

ish turns each answer into a position on a five-point scale with Semantic Similarity Rating (SSR), a published method (Maier et al., 2025). The answer is compared by meaning with first-person reference statements written for each point (for buying, from “There’s no chance I’d buy this” at the bottom to a clear yes at the top), across several sets of statements. The result for each participant is not one number but a spread: how much of their answer sits at each of the five points. Its mean falls between 1 and 5. For each iteration, the participants’ spreads are averaged into one, with its own mean, and iterations are ranked by that mean. The mean also gets a plain label: clearly wouldn’t, leaning against, divided, leaning toward, or clearly would. A participant with no answer to the question (someone who stopped early, for example) is scored from their end-of-session summary instead. Working out the read uses no generation, only the comparison by meaning, so it costs no credits and gives the same result each time for the same answers.

​

Two numbers for one read

The same read is shown two ways, and it helps to know which one you are looking at.

• The 1 to 5 mean. The raw result for each iteration, with its spread, band, and rank. This is what the CLI’s analysis results carry and what a benchmark comparison returns.
• The 0 to 100 index. The web app’s headline. It starts at 50 (balanced) and moves by half the difference between the share in the top two points and the share in the bottom two. It pools every iteration, weighted by how many people ran each.

Percentages in the breakdown are shares of the pooled spread, not counts of people, so a figure like 4% can appear in a panel of ten.

​

How many people a read needs

The web app shows no spread below 8 answers, and marks a read from 8 to 14 answers as an early read. From 15 it is shown as a firm read. That is also the sample size the run page suggests for a study that measures conversion intent.

​

What it does not claim

• It is not a prediction of how many people will convert, and not a sales, sign-up, or revenue forecast. “12 of 20 would buy” is not a reading of it.
• It does not compare across different studies or over time. It is most useful between iterations of one study, where the panel and questions are the same.
• The shape of the spread carries more than any single number.
• The method’s published reliability describes the method on labelled data, not ish’s own output. ish checked purchase intent against labelled reviews; sign-up and subscribe intent have not been checked separately against real outcomes.

When a study also uses the open gate, only the people who opened the content are asked, so the read describes openers.

​

Where it is set and read

Conversion intent applies to interactive and media studies. The web app does not offer it for chat studies. In a benchmark comparison, the purchase_intent field holds the 1 to 5 mean of the study’s first conversion event, whichever event that is. See benchmarks. Values are not validated when you set them: a misspelt event such as purchase-intent is stored and produces no read. Use the values in the table above exactly.

​

Related