← All posts

Playbook

How AI assistants decide what to recommend

A practical mental model for how ChatGPT, Claude, Gemini and Perplexity build a shortlist — and the five levers that decide whether your startup makes it on.

Cypress Team 2 min read

Quick answer

AI assistants build a recommendation by synthesizing patterns across the web, not by looking up a ranked database. Five levers decide whether you make the shortlist: source presence, category fit, endorsement strength, specific comparable claims, and freshness. You can't control the output, but you can systematically improve those inputs.


When a buyer asks an assistant “what are the best tools for onboarding analytics?”, the answer feels effortless — a tidy shortlist with a sentence of reasoning each. Behind that simplicity is a process you can reason about, and once you can reason about it, you can influence it.

Here’s a working mental model and the five levers that matter most.

The shortlist is a synthesis, not a lookup

The assistant isn’t reading a ranked database of tools. It’s drawing on patterns learned from a vast amount of text — reviews, comparisons, docs, discussions, editorial — plus, in some products, live retrieval from the web at answer time. It compresses all of that into a confident, compact recommendation.

Two consequences follow. First, consensus wins: the more independent sources point the same way, the more confidently the model names you. Second, the output is probabilistic: phrasing the question differently, or simply asking again, can reshuffle the list. Any honest measurement has to account for that variance.

The five levers

1. Source presence

Models recommend what they’ve “read about” repeatedly. If your product barely appears in the third-party sources for your category — review directories, “best X tools” roundups, community threads, comparison posts — you start with almost no signal. This is the single biggest lever for most startups.

2. Category fit

Assistants reason in categories. If buyers and writers consistently file you under “product analytics,” you’ll surface for product-analytics questions. If your positioning straddles three categories, the model has to guess which one you belong to — and often guesses wrong, or omits you to stay safe.

3. Endorsement strength

There’s a wide gap between “X is an option” and “X is the one most teams pick.” Sentiment and framing in your source material carry through. Lukewarm coverage produces lukewarm recommendations.

4. Specific, comparable claims

Models love attributes they can line up: who it’s for, what it integrates with, how it’s priced, what it’s best at. Vague differentiation (“powerful and easy to use”) gets dropped; concrete, comparable claims survive into the answer.

5. Freshness

The web’s consensus moves. New roundups, launches, and discussions can shift a shortlist within weeks. Tools that keep earning fresh, corroborating coverage hold their position; tools that go quiet drift off.

Turning the model into a checklist

For any buyer question you care about, ask:

Answering those questions by hand, across dozens of prompts and four assistants, is tedious and inconsistent. That’s the measurement problem Cypress automates: it runs the prompts, scores appearance and endorsement strength, surfaces the competitors winning each question, and traces the sources behind the answers — so the five levers become a prioritized to-do list instead of a hunch.

You can’t control what an assistant says. But you can systematically improve the inputs it reads — and that is what moves the shortlist.

Frequently asked questions

How does ChatGPT decide which products to recommend?

It synthesizes patterns from a large body of text — reviews, comparisons, docs, discussions, editorial — plus, in some products, live web retrieval. The more independent sources point the same way, the more confidently it names you. The result is probabilistic, so it changes across runs.

Why does the same question give different recommendations each time?

AI outputs are probabilistic. Re-phrasing the question, or simply asking again, can reshuffle the shortlist. That's why visibility has to be measured as a distribution across many runs, not from a single check.

What is the biggest factor in AI recommendations?

For most startups it's source presence — how often and how positively you appear in the third-party sources models read. If you barely show up in your category's reviews, roundups and discussions, you start with almost no signal.

Can I directly control what an AI says about my product?

No. You can't edit the model's output. What you can control are the inputs it reads — the sources, the clarity of your positioning, and the specificity of your claims — which is what moves the shortlist over time.

Keep reading


See it for your startup

Find out if AI assistants recommend you

Cypress benchmarks real buyer prompts across ChatGPT, Claude, Gemini and Perplexity and scores how strongly you're recommended. Free plan, no credit card required.

Start free