Limitations and accuracy
What a Peach report can and cannot tell you. A sample is not a census, and an API answer is not the app.
Updated
In short. Peach measures a small sample: one answer per question per engine, asked through APIs without web search. It is good evidence of what a model says when asked cold. It is not a measurement of what every buyer sees.
A sample is not a census
A check collects at most sixteen model answers. Your mention rate is a fraction of those.
- One answer more or less moves the overall figure by about six points.
- Two checks of the same site on the same day can differ.
- A 0% result means you were not named in this sample. It does not prove you are never named.
Use the number to compare yourself with competitors inside one report, and to see direction over several reports. Do not present one figure as a precise measurement.
Answers vary
The same question can produce different answers depending on:
- Who asks. Assistant apps personalise.
- Where they ask from. Apps vary by country.
- How they phrase it. A small change in wording changes the list.
- When they ask. Models are updated.
Peach asks each question once, in one wording.
An API answer is not the app
ChatGPT and Gemini are asked through their APIs, with no web search. The apps can search the web. So the report can show you missing where the app would have found you through a search, and the reverse.
If your own test in the ChatGPT app disagrees with the report, both can be right. They are different measurements.
Google Search is not AI Overview
The Google check reads the answer box and top results of an ordinary search. Most questions have no answer box, so "No answer" is the usual Google verdict. Peach has never read an AI Overview.
Matching is by name
A mention is detected by finding your name in the text.
- A brand that is also an ordinary word is harder to match correctly. Peach is cautious with these.
- An answer that describes you without naming you is not a mention.
- An answer that names you unfavourably still counts as named.
Competitor lists are not fixed
The competitor list is built by a model and a web search, and can change between checks. A wrong list makes every comparison wrong. Check it, and correct it in Brand Hub.
Market detection is conservative
Peach names a country only on evidence and otherwise says global. It reads your site as served to its servers, so a site that shows different content by visitor location may be read differently from how your buyers see it. Set the market yourself if it matters. See Market and geography.
Citations are domains in text
The Citations tab counts domains written in answers. They are not retrieved sources, and only some domain endings are counted. See Citations.
Only three gaps are listed
A report lists up to three missed questions, even when more were missed.
Query Fan-Out is an estimate
The related questions are generated by a model. Peach does not observe what an assistant searches for.
Generated articles need a reader
Quality checks count words and match phrases. They do not judge accuracy or sense. Web research fails for some sections, and the report says which. Read an article before you publish it.
What Peach does not do today
Coming soon. Scheduled re-runs and trend alerts, Perplexity, PDF export, shareable report links, updating a published post in place, and a comparison article format.
If something in a report looks wrong, email hello@gotopeach.com with the website address. Wrong results are the most useful thing you can send.