How we measure

The engines

We ask three engines, each through its API with web search turned on. The report uses these labels:

OpenAI (GPT via API, web search)

OpenAI's Responses API with its web search tool.

Perplexity (Agent API, web search)

Perplexity's Agent API with its web search tool, with the model pinned to perplexity/sonar.

Claude (API, web search)

Anthropic's API with Claude's web search tool.

Every answer is stored with the model ID, the search settings, the location setting and the time it was asked, and the report prints them.

Gemini isn't included. Google's terms for Gemini answers grounded in Google Search don't allow analyzing or reselling them.

The questions

We write 20 questions across five kinds: best for, use case, comparison, price limit and alternatives. They come from your store's category and the competitors you list.

Jake edits the draft, then the set is frozen and reused for every later run of your store. The report lists all 20.

Three runs

AI answers vary from one ask to the next. Each question runs three times on each engine, up to 180 answers in all, so one lucky or unlucky answer can't decide your result.

Here is how those answers line up in a report. Each row is a question and each cell is one answer; choose a brand to see which answers named it.

What counts as named

A brand is named when an answer recommends it or lists it as an option to buy. A brand mentioned only to warn against it is a negative mention, not named, and the report shows confirmed warnings on their own line.

A product counts for the brand that makes it. Retailers and marketplaces such as Amazon, REI and Walmart are listed under where AI sends shoppers to buy, not as competitors.

An extraction model reads each answer. A report ships only when that reader scores at least 95% against answers we labeled by hand, both on the brands it finds and on the ones it misses. Jake then confirms every brand name in your report.

Completed answers

An answer is completed when the engine returned text and the reader produced a valid record from it. Answers that fail are left out of the count and listed with the reason.

An engine is included when at least 48 of its 60 answers completed. A report needs OpenAI plus at least one of Perplexity and Claude. If a run falls short we run it again, and if the second run falls short too, you get a full refund.

Share of answers

Share of answers is the number of completed answers that named a brand, divided by all completed answers. Every answer counts the same, whichever engine gave it.

The leaderboard shows the share with the count beside it and splits the count by engine. Prompt coverage, shown per engine, counts the questions where a brand was named in at least 2 of that engine's 3 runs.

Citations

Sources come from each engine's structured citation data, not from links written into the answer text. A site counts as cited when it's in an answer's citation list.

Sites are grouped by domain, so every page on reddit.com counts toward reddit.com, and the report lets you open the pages underneath.

The ChatGPT app screenshots

Each $99 report includes three screenshots from the ChatGPT app, for the questions where you were named least and competitors were named most.

Jake takes them by hand in a temporary chat with memory off. Each caption gives the date, the app, the account state and the location. A screenshot is never retaken or swapped because of what it shows. They're examples to compare against: every count in the report comes from the API answers.

What the report doesn't claim

It doesn't say what every shopper sees in consumer apps, which can differ by account, history and location. It doesn't measure traffic or sales.

It's dated, because answers change as models and the web change. A later run of the same questions shows what moved.

Who can see your report

Your report lives at a private link with a long random key. It isn't listed anywhere, search engines are told not to index it, and the key isn't passed on to the sites you click through to.

Screenshots are stored privately and shown through short-lived links. The link keeps working after a refund.

More in our privacy policy.