Outpick
Recommendation intelligence

AI engines know your brand. Do they recommend it?

Outpick measures how often AI answer engines recommend you — with intervals, evidence and a method you can audit. Not a score. A report you can act on.

No login. One field. Returns a real range and its caveat.

01

The gap

71%
Named when the category is asked about

The engines know who you are. Every tool in this category stops here and calls it visibility.

47–62%
Actually recommended to the buyer

The range, not a point. This is the number that decides whether someone buys from you.

The distance between being known and being chosen is the only measurement that pays for itself.

Illustrative figures · 412 eligible observations · 32 buyer-journey prompts · trail running shoes, Netherlands. Your own numbers will differ.

02

Why your rank tracker can’t see this

You rank #1 on Google and you don’t exist in ChatGPT.

A search ranking

Ten blue links. The buyer chooses among them, and you can see your position.

An AI answer

One paragraph. Two or three brands named. The engine has already chosen for the buyer.

What nobody measures

Whether you were one of them — and, when you weren’t, which source made the case for someone else.

03

What you actually get

A document, not a dashboard. Here is a page of one, at full size.

Read the whole sample
Outpick · Kestrel · AI Recommendation Report
02

Recommendation probability

Recommendation probability by engine

A recommendation is counted only when the brand is positively proposed for the buyer’s stated need.

ChatGPT61%
Gemini56%
Copilot49%
Perplexity44%

Based on 412 eligible observations across 32 buyer-journey prompts, at least 7 repeated runs per engine, 3–16 Aug 2026, with consistent behaviour across repeated runs. The range shows where the true rate is likely to sit at this sample size.

Illustrative sample data — a fictional brand in a real category. Figures demonstrate the method, not a live measurement.

04

How it works

01ScopeWe agree the category, market and the competitors you are actually measured against. Nothing is inferred from your website.
02Prompt portfolioThirty-odd questions a real buyer asks — category recommendations, head-to-head comparisons, alternatives. You approve it before anything runs.
03Repeated scansEach prompt runs at least seven times per engine, because answer engines are not deterministic. One run is an anecdote.
04ReportA document your VP can read in thirty seconds and your agency can work from for a quarter. Ranges, denominators, sources, evidence.
05Corrective actionsRanked by expected effect on recommendation rate, each tied to the observation that justifies it. Usually fewer than five things matter.
05

The measurement contract

We publish ranges, never a single number.

Wilson intervals at 95%, so the figure states its own uncertainty instead of hiding it.

Every figure carries its denominator.

“47 of 88 eligible observations”, on the same line as the percentage. Always.

Unavailable is not zero.

When an engine cannot be measured we say so. Recording a zero invents a decline that never happened.

There is no Outpick score.

A weighted composite cannot be falsified or reproduced. We will not ship one — this is permanent.

Design partner programme

One report, operator-delivered, €499.

Ten to fifteen slots, then the programme closes. You get the full report, a walkthrough, and direct influence on what the product measures next.