How Optimz measures AI visibility

Optimz measures ChatGPT, Claude, Gemini and Perplexity for every client, through each engine's API with web search turned on. Grok, DeepSeek and Meta AI can be added per client; DeepSeek and Meta AI search through an added web-search tool, so their sources can overlap with Perplexity's.

Google AI Overviews and AI Mode are read from Google results pages. The ChatGPT and Gemini consumer apps can also be checked as a signed-out user sees them, which shows how far the API answers are from what people see.

Each tracked question is asked once a week on each API engine by default. An agency can raise this to three per client, and those samples are spread across the week rather than run at the same moment, because answers from the same instant are correlated. Google AI Overviews, AI Mode and the consumer apps are read once a week per question: buying the same results page again adds cost, not new information.

• Mention rate: the share of answers that name the client.
• Recommendation rate: the share of answers that suggest the client for what was asked, not just name it.
• Citation coverage: the share of answers that link to at least one of the client's own pages as a source.
• Share of voice: the client's mentions as a share of all mentions of the client and its tracked competitors in the same answers.
• Position: where the client appears in the answer's list, shown as detail only, because order changes from run to run.

Whether an answer names, recommends or cites the client is read by a parser that also checks the brand isn't a different company with a similar name. Every rate links to the answers behind it, so any number can be checked by reading them.

Every rate is shown with the number of answers it's based on (n) and a 95% Wilson confidence range. The range is the honest width of the estimate: at n = 100 and a rate near 50%, it runs from about 40% to 60%.

Rates are pooled across questions, runs and weeks, because single answers are noise: independent 2026 studies found two runs of the same prompt share only about 40–60% of the brands they name.

A question and engine with fewer than 20 answers shows “Collecting” instead of a percentage, because its range would be too wide to read. A question counts as a competitor win only with at least 12 pooled answers and a competitor's range clear of the client's, so an action never fires on a handful of answers.

Headline rates cover the last 4 weeks and are compared with the 4 weeks before, with the single week kept as detail. Weekly numbers at typical volumes move by 5–10 points from noise alone; a rolling window shows the trend without hiding it. Model changes reported by an engine are marked on the charts.

A change between two periods is tested directly on the difference, not by checking whether two ranges overlap (overlap is far too strict and makes real changes look like noise). The test treats answers to the same question as a group, since they move together, and a change is called real at the 5% level. Anything smaller is shown as “no clear change”.

When a fix ships, Optimz compares the questions it targets before and after against the client's other questions on the same engines over the same weeks. The net effect is the change in the treated questions minus the change in the control questions, with its range. Without a control, before/after numbers mostly measure platform-wide drift: in an Ahrefs schema test, a +43% raw change on AI Mode was about +2% net.

Before a fix ships, Optimz estimates the smallest change the test can detect (at 80% power) and by when, so “no change yet” is a planned outcome, not a surprise. Earlier signals are shown in order: the page changed, an AI crawler fetched it, an own page was cited, then the mention rate moved.

Every answer is labelled with how it was collected: “API with web search”, “API with added web search”, “Google results page” or “Consumer app”. API answers are not what a person sees in the app: published 2026 tests found the API and the ChatGPT app share only about 20–30% of the brands they name. So the labels are never hidden and the two are never mixed silently.

Answers are collected signed out, with no account history or memory. They are an unpersonalized baseline: a person's own history can shift answers by a few points, which no outside tool can see.