AI search performance

What does an AI search performance report look like?

The strip at the top of this page is ours. It is attensira.com's own workspace for the 30 days from 5 September to 4 October 2026, read over the Attensira MCP on 4 October: 61 questions a buyer of an AI SEO product asks, read on ChatGPT and Perplexity, 4,524 answers in all.

Read the sample sizes first. Each rate sits on its own n, and the two engines are reported apart because they do not behave alike: our share of voice is 4.21% on ChatGPT over 2,349 answers and 4.51% on Perplexity over 2,175. 105 reads failed in the window; they are left out of n, not counted as answers that ignored us.

Then read the headline with suspicion. A blended 4.35% looks like a modest start, but it is an average of two very different groups of questions. Split by whether the question contains the brand name, and almost all of it comes from the branded ones; that split is laid out on AI brand mentions. A report that does not separate them flatters you.

An honest report is mostly unflattering at the start. That is what makes the next reading worth anything.

What does AI search performance measure?

Four things, which move for different reasons and must not be blended:

1. Mentions. The share of answers that name your brand. See AI brand mentions and share of voice.
2. Citations. The share of answers that link a page on your domain. See source citation.
3. Sources. Which domains the answers lean on for your questions, including ones that are not yours. These tell you where to be present.
4. Access. Which AI crawlers fetched your pages. A page that is never fetched cannot be read at answer time.

The first three are samples. The fourth, from your server logs, is a census: every fetch by a named agent appears with its user agent.

Why is it measured by sampling?

Because no AI engine gives you an impression log. ChatGPT does not tell you when it mentioned you. The only way to know is to ask the question and read the answer, many times, on each engine, in each country you sell in.

That has three consequences:

• Every rate needs its n. Attensira returns every rate as a value with the number of answers behind it, and below 20 answers the dashboard shows the count instead of a percentage. See Reading the numbers.
• Not measured is not zero. An engine nobody asked, or a day with no readings, is a dash, not a 0. A measured zero means the answers were read and nobody named you.
• A change has to clear the noise. Attensira tests each change with a two-proportion test at 95% confidence. A movement that does not clear it is reported as "nothing we can prove yet", not as a win.

Can Search Console show AI search performance?

For Google's own AI surfaces, partly. Google states that traffic from AI features sits inside the ordinary Web search totals in Search Console, and its help centre says that as of 31 August 2026 a generative AI report with impressions for AI Overviews and AI Mode is available to all websites. See Google Search Console.

For everything outside Google, no. ChatGPT, Perplexity and Claude report nothing to Search Console. A full picture puts Search Console beside sampled answers and crawler logs, read on the same days.

What are the objections to measuring AI search performance this way?

"Give me one score." A blended score hides which of the four things moved, and hides the sample size. Attensira never collapses mentions and citations into one number. See AI visibility score.

"Our number jumped this week." Check n, then check whether the change cleared the test. Small samples move in large steps on their own.

"Show me the revenue." No one can join an answer to a later sale. Attensira shows what changed in the answers and in your traffic beside the work it shipped, and does not claim causation.

"We'll measure once a quarter." Engines change week to week. Daily readings give a baseline you can compare a change against.

How does Attensira measure and improve it?

Attensira reads your questions on the engines every day, stores every answer verbatim, and reports named and linked rates per engine with their sample sizes, plus the sources that decided each answer. Its crew then does the work to move those numbers, ships each change after your approval, and reads the answers before and after. A report puts readings like the ones above, for your own brand, next to the work that was shipped in the same window.