Method

CiteSieve Methodology: How AI Visibility Is Measured

CiteSieve measures AI visibility by asking each tracked question once a day on every engine switched on, storing each answer in full, and counting the share of answers that name your brand. Every figure is shown with the number of answers behind it and a 95% Wilson interval, and week-on-week changes are tested for significance.

Key points

  • Each question is asked once a day on each engine, and every answer is stored in full with its sources.
  • Your brand and competitors are matched by exact text, not by a model's judgement.
  • Every figure shows its sample size and a 95% interval, because AI answers vary between runs.
  • A week-on-week change is labelled significant only when it passes a test, and needs at least 20 answers in each week.
  • The method measures what engines answer to your questions, not how many people saw those answers.

How does CiteSieve collect AI answers?

CiteSieve asks each tracked question once a day on every engine you have switched on, and stores the answer it gets back. The engines are ChatGPT, Google AI Mode, Gemini, Perplexity, Copilot and Google AI Overviews.

Answers are collected through a third-party collection service, as a searcher in your chosen country would get them. They are not tied to an account of yours, so there is no chat history or personal memory behind them. A question can be up to 32 words long.

  1. You add your brand, your competitors and the questions your customers ask.
  2. Each day, every question is sent to each engine you have switched on.
  3. Each answer is stored in full with the time, the engine, a SHA-256 fingerprint of the text and the sources the answer cited.
  4. The brands the answer names are listed in the order they appear.
  5. The figures on your dashboard are recalculated from the stored answers.

An engine that returns no answer for a question that day is not counted and not charged.

How are brands and sources read from an answer?

A language model reads each stored answer and lists every brand it names, in order. Your own brand and the competitors you entered are then matched by exact text, so whether you were named never rests on a model's guess.

Brands you did not enter still count towards share of voice. Sources are the links the answer itself cited. Tracking parameters are removed from those links before they are stored, so the same page is counted once.

What do visibility and share of voice mean in CiteSieve?

Visibility is the share of stored answers that name your brand. Share of voice is your share of all brand mentions in those answers. Both can be split by engine and by question.

FigureHow it is calculatedShown with
VisibilityAnswers naming your brand, divided by all answers collectedNumber of answers and a 95% interval
Share of voiceMentions of one brand, divided by mentions of every brand namedNumber of answers
Week-on-week changeVisibility in the last 7 days minus the 7 days beforeA label: significant, within noise, or not enough data yet
SourcesDomains the answers cited, counted across answersNumber of answers citing each

For the share of voice formula and a worked example, see the guide to AI share of voice.

Why does every figure carry a sample size and an interval?

AI answers change from one run to the next, so a visibility figure is a sample, not a census. OpenAI's own documentation describes model output as non-deterministic by default. CiteSieve therefore shows how many answers stand behind each figure and a 95% Wilson interval around it.

The table shows how the interval narrows as answers accumulate. These are our own calculations for an observed visibility of 30%, using the Wilson method.

Answers collected95% interval around 30%Width
3016.7% to 47.9%31.2 points
9021.5% to 40.1%18.6 points
18023.8% to 37.1%13.3 points
36025.5% to 34.9%9.4 points

One question on three engines gives about 90 answers in 30 days. A figure built on 30 answers is a rough reading. A figure built on a few hundred is one you can act on.

How does CiteSieve tell a real change from noise?

CiteSieve compares the last 7 days with the 7 days before and tests whether the difference is larger than run-to-run variation would explain. The change is labelled significant only when it passes that test.

  • Significant rise or fall: the difference is unlikely to be chance at the sample sizes collected.
  • Within noise: the difference could be ordinary variation between runs.
  • Not enough data yet: either week has fewer than 20 answers, so no verdict is given.

The background to this test is in how accurate AI visibility trackers are.

What are the limits of this method?

The method measures what the engines answer to a fixed set of questions. It does not measure how many people asked those questions or saw the answers.

  • One answer per engine per question per day. Samples build up over days, so early figures have wide intervals.
  • No personal context. A signed-in user with history or memory may see a different answer.
  • Country level. Answers are collected for a country, not a city.
  • Engines change. Models and their search behaviour are updated over time, which can move figures for reasons unrelated to your site.
  • Cited sources only. CiteSieve records the links an answer shows, not every page the engine read.
  • Your question set matters. Visibility on your questions says nothing about questions you did not track.

How can you check the numbers yourself?

Every stored answer can be opened and read in full, and all of them can be exported as CSV. Each answer carries its engine, timestamp and SHA-256 fingerprint, so you can recount any figure from the raw text.

Stored answers are kept for up to two years. The plans and what a credit buys are on the pricing page, and the tool comparison sets CiteSieve beside other trackers.

How to measure whether it worked

Use the same fixed questions for at least four weeks before judging a change you made to your site or your coverage elsewhere. Read the interval, not only the headline figure, and wait for a change labelled significant before acting on it.

Compare engines separately. A brand can rise in ChatGPT and stay flat in Google AI Mode, because the two draw on different sources.

Questions and answers

Does CiteSieve use the engines' APIs or their consumer products?

CiteSieve collects answers through a third-party collection service, as a searcher in your chosen country would get them from each engine. It does not use an account of yours, so the answers carry no chat history or personal memory. An answer collected this way can differ from what a signed-in user sees.

How many answers do I need before I can trust a figure?

It depends on how precise you need to be. By our calculation, an observed 30% visibility has a 95% interval about 31 points wide at 30 answers and about 9 points wide at 360 answers. CiteSieve shows the interval on every figure so you can judge it directly.

What happens when an engine gives no answer?

An engine that returns no answer for a question on a given day is left out of the count for that day and is not charged. The sample size shown beside each figure only includes answers that were actually collected and stored.

Can I export the raw answers?

Yes. Every stored answer can be read in full inside CiteSieve and all of them can be exported as CSV, with the engine, the time, the brands named and the sources cited. That lets you recount any figure or analyse the answers in your own tools.

Sources

  1. OpenAI API docs: Advanced usage (reproducible outputs)
  2. NIST/SEMATECH e-Handbook: Confidence intervals for a proportion (Wilson method)
  3. NIST/SEMATECH e-Handbook: Do two processes have the same proportion of defectives?
  4. SparkToro and Gumshoe: AIs are highly inconsistent when recommending brands or products

Sources were read on 12 November 2026. Engines and their documentation change; check the linked page for the current wording.

More guides

Find out what AI says about your brand.

A site audit on the page in seconds, then live answers by email. No account, no card.