How we measure

• Your prompts, asked in your workspace's country and language, without being signed in to any account: no memory, history or personalization.
• Visibility is the share of answers in which your brand appears: the answer names you, or cites a page of your website as a source.
• Each AI has its own figure. The total is the average of the AIs with data, and every AI counts the same however many answers it gives.
• Empty answers and failed readings do not count: they neither add to nor subtract from your visibility.

ChatGPT

On chatgpt.com, the way a person sees it, asking it to search the web on every question. If a reading fails, the API answer is kept apart as another source and does not mix with the chatgpt.com answers in the figure.

AI Overviews

The AI summary at the top of Google's results, with the sources it shows. When Google shows no AI Overview for a prompt that day, the answer counts as empty and does not enter the average.

Perplexity

Its Sonar API through OpenRouter, with the web search Perplexity runs on its own.

Claude

Its model through OpenRouter, searching the web with Parallel, a search provider chosen by Placton, not the one on its own website (there Claude uses Brave). Its answers can differ from what a person sees on claude.ai.

Grok

Its model through OpenRouter, searching the web with Parallel, a search provider chosen by Placton, not the one on its own website (there Grok uses Google and X). Its answers can differ from what a person sees on grok.com.

How often

ChatGPT and AI Overviews every day, wherever they can be read that way; Claude, Perplexity and Grok every week. During the free trial, the weekly AIs are measured once, when you sign up.

• Cited: a source that comes with the answer. Only these count as citations, and only these turn a citation of your website into an appearance. What that means on each AI:

ChatGPT

The sources the answer shows on chatgpt.com. The pages its search only returned are kept apart and do not count.

AI Overviews

The sources Google shows with the AI Overview.

Perplexity

The sources Perplexity lists with its answer.

Claude · Grok

The sources OpenRouter returns with the answer. Their APIs do not say which of them a person would see on claude.ai or grok.com.

• Returned: a page that ChatGPT's search got back while it prepared the answer. We keep them apart: coming up in that search does not mean ChatGPT read the page, only that it was among the results it received. Only your server logs can show a real visit.

• The margin is worked out within your prompts. Your prompts are fixed, so the uncertainty comes from how each answer changes from one run to the next. It does not say what other prompts would give.
• How much each prompt varies comes from its last 8 weeks with the same source, pulled a little towards the average of all your prompts.
• Answers from consecutive days that look alike count for less: if an AI barely changes from one day to the next, 7 daily answers are worth not much more than one. We measure that for each AI every day with the last 8 weeks of data. Until there is enough data we use prudent values (for AI Overviews, that it barely changes from one day to the next).
• In 30 and 90-day views the figure itself does not change: the margin goes around the same share of answers as always.
• In 7-day views, the AIs tracked every week use their last 4 weeks, because a single weekly run is too little. That is why the 7-day figure can differ a little from the one before this change.
• On a single prompt we show the count ("5 of 7") and no margin: one prompt on its own varies a lot, so decide with the total.

Method: Wilson interval with the effective sample size of each AI; design effect from the day-to-day autocorrelation; the total combines the interval of each AI (MOVER, Zou and Donner).

• The visibility arrow compares each prompt with itself in the two periods, only on prompts with answers in both. What makes one prompt different from another cancels out, and only the noise within each prompt is left.
• It says “real change” when the change is beyond the normal variation of the answers (95% confidence) and “within noise” when it isn't. Its help text also gives the smallest change the data can detect.
• 7 days: the daily AIs compare the last 7 days with the 7 before; the weekly AIs, their last 4 weeks with the 4 before. 30 and 90 days: each period with the one before.
• An AI is only compared when the two periods share the same source. When the source changes (for example, ChatGPT moving to chatgpt.com), that AI is left out of the comparison and the day is marked on the chart.
• The sentiment, position and mentions arrows have no test yet: they are shown in grey, as a rough guide. Average position moves a lot because AI answers rarely repeat the same order (SparkToro and SISTRIX).

An alert fires only when all of this holds:

• The change passes a test within your prompts, with false alarms kept in check across all the tests of the same reading (Benjamini-Hochberg, q = 0.05).
• It is 5 points or more.
• At least 3 prompts move: a single prompt does not make an alert.
• Daily alerts must show up two days in a row.

• Daily reading (ChatGPT and AI Overviews): the last 7 days against the 28 before, for your visibility on each one and for each tracked competitor's lead over you.
• Weekly reading, with the summary email: your total visibility, each AI (the weekly ones, their last 4 runs against the 4 before) and each competitor's lead.
• The same alert does not repeat within 14 days. Daily alerts start with 14 days of data; those of Claude, Perplexity and Grok, after 8 weekly runs.
• Alerts are part of paid plans, and you get at most one alert email a day.

Visibility carries one of two labels. The thresholds are our own, public, and based on the IAB framework (query volume and sample size), with precision added.

Said plainly: with 20 prompts your figures are directional. They show the trend; to decide with a figure on its own you need more prompts.

• We ask without being signed in. Someone signed in, with memory or history, may get different answers.
• One country and one language per workspace.
• Claude and Grok search the web through OpenRouter, not with the search engine of their own websites.
• AI Overviews only count on the days Google shows one.
• We see what the AIs answer, not what your customers then click or buy.
• What we keep: the text of every answer, with the brands and sources in it, for as long as you have an account. What else arrives with the ChatGPT and Google answers (the pages ChatGPT's search returned, Google's results for the same query) is kept in detail for 90 days; after that, only the counts.

Every change in what we ask the AIs or in how we count, with its date. In your dashboard, the changes that affect your period also appear under the trend chart.

1. October 30, 2026 · Calculation · Recalculated back in time

   Visibility shows its margin of error

   Visibility now comes with its 95% margin, worked out within your prompts, and a label: directional or decision-grade. In 30 and 90-day views it is the same figure as before. In 7-day views, the AIs tracked every week (Claude, Perplexity and Grok) use their last 4 weeks instead of their last run, so that figure can move a little from this day; the chart was recalculated back in time with that rule.

2. October 30, 2026 · Calculation · Recalculated back in time

   Changes compare each prompt with itself

   The visibility arrow compares each prompt with itself in the two periods, only on prompts with answers in both, and says whether the change is real or within the noise. The sentiment, position and mentions arrows have no test yet and are shown in grey.

3. October 30, 2026 · Calculation · Applies from that day

   Alerts only fire when the change is beyond the normal variation of the answers

   Before, alerts fired with fixed thresholds of 10 and 5 points. Now each change is tested within your prompts, false alarms are kept in check across all the tests of the day, and daily alerts must show up two days in a row. Alerts already sent stay as they were.

4. October 30, 2026 · Calculation · Applies from that day

   Pages that ChatGPT's search only returned no longer count as citations

   Before, when an answer on chatgpt.com showed no sources (about 1 in 200 answers), the first 10 pages its search returned counted as citations, and through them as indirect mentions. From this day only the sources the answer shows count. Earlier answers stay as they were; if they are ever corrected, it will be added to this log.

5. October 19, 2026 · Measurement · Marked on the chart

   ChatGPT is read on chatgpt.com, the way a person sees it

   Before, ChatGPT was read through its API. Where chatgpt.com can be read, it is now read there every day. If a reading on chatgpt.com fails, the API answer is kept as a separate source and does not count in the figure.