How we measure

What we measure

For every question and every engine, each morning, we record four things:

• Named. Did the answer mention your business by name?
• Address given. Did it give your web address, as a link or written out? Being named without an address still leaves the person to search for you.
• Recommended instead. Which other businesses did the answer name, in the order and words the engine used.
• Pages read. The web pages the engine says it used to write the answer.

The full answer is kept word for word next to those four fields, so every number on a report can be traced back to a sentence an AI actually wrote.

The questions

We read your website and write the questions a real customer would ask. Real questions to AI assistants are long, start from a problem, and often say "I" or "my": "I run a small race and need live tracking without timing chips", not "race tracking software". We write them that way.

You see every question before anything runs, and you can edit or replace any of them. We start with ten. A few questions that matter tell you more than hundreds that don't, and the quiet ones can be swapped out.

How we ask

Each question goes to ChatGPT, Claude, Gemini and Perplexity through the official developer API each company provides, with its web search tool switched on. We never automate the consumer apps. Their terms don't allow automated use like this, and an answer we couldn't get honestly isn't one we want to sell.

The questions are asked fresh each morning with no account history or saved memory attached, and with the same settings every day, so a change on your report reflects a change in the answers rather than in how we asked.

Why live search

A language model on its own only knows what was in its training data, which stops at a fixed date. A business that launched since then may not exist for it at all. With web search on, the engine looks up current pages before answering and tells you which ones it used.

That matters twice. It is how the engines answer most questions about products and services today: Claude searches when it judges a search relevant, and Gemini decides per question whether a search would improve the answer. And the list of pages read shows you why you were or weren't recommended, which points straight at the page to write or fix.

Why runs, not ranks

Ask an AI the same question twice and you will often get a different list. When SparkToro and Gumshoe ran 12 recommendation prompts 2,961 times, ChatGPT and Google's AI returned the same list of brands in under 1% of repeated runs, and the same list in the same order in under 0.1% (Search Engine Journal, January 2026). Even with randomness turned down to zero, the same prompt can give different outputs, because of how the servers batch requests (Thinking Machines, September 2025).

What does hold up is how often a name appears. So a report says "named in 4 of 5 runs", over the last five mornings, and shows the five results as a strip: . We never turn that into a single "#1". The reasons are set out in Why AI answers change every time you ask.

What this can't tell you

• It is not one person's screen. The consumer apps can use a person's history, memory and location. Our questions carry none of that, which makes them comparable day to day but not identical to any one user's view.
• A small number of runs is a small sample. "4 of 5" is a rate over five mornings, and we show it that way rather than as a percentage with false precision.
• The pages read are what the engine reports. An engine may read more than it lists. We show exactly what it returned.

If you want to try the manual version first, we wrote up how to check whether ChatGPT recommends your business by hand.