Methodology
How we measure whether AI picks you
Last updated 2026-10-15
What we ask
We write the questions a buyer would type: the best tool or app in your category, options for your audience or use case, alternatives to a named competitor, and one question about you by name. Unbranded questions measure whether you get picked. The branded question checks whether engines describe you correctly.
The category comes from you, or from your homepage when you leave it blank. Without a category we ask only the branded question, because buyer questions about a guessed category would mislead you.
The Pick Check that comes with a free account (one a day) asks 3 questions 3 times on each engine. Paid reports use a larger question set with the same method.
Where the answers come from
Every answer comes from the engine's official API with web search turned on. We never scrape consumer apps, which their terms forbid. We ask the question in one plain message, with no instructions, the way a person would.
| Engine | How we ask | Default model |
|---|---|---|
| ChatGPT | OpenAI Responses API with the web_search tool, via Vercel AI Gateway | openai/gpt-5.4-nano |
| Claude | Anthropic Messages API with the web_search tool, via Vercel AI Gateway | anthropic/claude-haiku-5.5 |
| Gemini | Gemini API with Grounding with Google Search, via Vercel AI Gateway | google/gemini-3.5-flash-lite |
| Perplexity | Perplexity Sonar (built-in search), via Vercel AI Gateway | perplexity/sonar |
| Grok | xAI Responses API with the web_search tool | grok-4 |
API answers can differ from what the consumer apps show, because the apps add personalization, memory and their own retrieval. Microsoft Copilot and Google AI Mode have no public API. We cover them indirectly through the Bing and Google indexes they read.
How we count a pick
An answer picks you when it names you as a whole word (so "Linear" does not match "linearly"), mentions your domain, or cites one of your pages. Failed calls are reported and left out of the rate.
Engines give different answers to the same question on different runs. A SparkToro and Gumshoe study of 2,961 prompts found less than a 1-in-100 chance of seeing the same list of brands twice. So we report a pick rate per engine with a 95% likely range (a Wilson interval), never a rank, and we never blend engines into one number.
"Sources cited instead" counts the domains engines cited in answers that did not pick you. Those are the pages that shaped the answer.
Crawler access and readiness
We fetch your homepage and key pages as each of 20 AI crawler tokens and read robots.txt the way RFC 9309 specifies. A firewall that verifies crawlers by IP can treat our test differently from the real bot, so when a normal browser fetch is also refused we report the result as unknown, not blocked.
The readiness score covers what is on your own pages: crawlability, a clear description, evidence, entity markup, buyer-question pages and comparison pages. It is reported separately from the live AI answers, because what other sites say about you matters more than anything on your own site.
Verifying fixes
After you ship a fix, a re-run crawls the page to confirm the change is live, then asks the same questions the same number of times on the same engines. Only then do we compare before and after.
Index coverage
An engine can only pick a page that is in the index it searches: Google's for Gemini and AI Overviews, Bing's for Copilot, Brave's for Claude. For monitored sites we read the sitemap, pick up to 25 key pages (homepage, pricing, comparisons, features, docs, then the newest) and check each one every week.
- Google: Search Console's URL Inspection verdict, through read-only access you grant. We never change your Search Console.
- Bing: crawl data from Bing Webmaster Tools with your API key. Bing reports crawls, not an index verdict, so we count a page crawled with HTTP 200 as indexed.
- Brave: Brave has no index API, so we search your site in Brave Search and look for the exact URL.
- IndexNow: once your key file is live, we notify Bing, Yandex, Naver, Seznam and Yep of new and changed pages. Google does not use IndexNow.
What AI cites data
For our category data pages we ask the same kind of buyer questions without naming any brand, three times on each engine, every month. We count the share of answers citing each site at least once, and the share naming each brand on a fixed list per category. We publish only totals, never answer text, and only when at least two engines and 30 answers came back.
What we do not do
- Guarantee rankings or recommendations.
- Post on Reddit, forums or review sites as you, buy votes, or run fake accounts.
- Sell llms.txt as a fix. Large studies found no effect on citations.