AI answers change every time you ask. Measure them like it.

Citebell tracks how often AI answer engines mention your clients' brands, puts an error bar on every number, and re-runs the prompts to check that a fix worked. Built for SEO and marketing agencies.

Measured on ChatGPT Claude Perplexity Google AI Overviews and AI Mode
Start free trial

14 days · up to 20 prompts · no card. Measurements start as each engine goes live.

Tracked prompt ChatGPT measured via API

Illustrative data

“best payroll software for a 20-person company”

Asked 20 times. Does the answer name Northwind Payroll?

  1. Answer 1: names Northwind Payroll
  2. Answer 2: names Northwind Payroll
  3. Answer 3: doesn't name it
  4. Answer 4: names Northwind Payroll
  5. Answer 5: names Northwind Payroll
  6. Answer 6: doesn't name it
  7. Answer 7: names Northwind Payroll
  8. Answer 8: names Northwind Payroll
  9. Answer 9: names Northwind Payroll
  10. Answer 10: doesn't name it
  11. Answer 11: names Northwind Payroll
  12. Answer 12: doesn't name it
  13. Answer 13: names Northwind Payroll
  14. Answer 14: names Northwind Payroll
  15. Answer 15: doesn't name it
  16. Answer 16: names Northwind Payroll
  17. Answer 17: doesn't name it
  18. Answer 18: names Northwind Payroll
  19. Answer 19: doesn't name it
  20. Answer 20: names Northwind Payroll

13 of 20 answers name it

95% CI 43–82% · n = 20

Too wide to quote as one number, so Citebell reports the band the estimate falls in, usually (50–80%), with the interval beside it.

rarely sometimes usually almost always
  1. Run 1: 65%, interval 43–82%

Each run lands somewhere different. The interval says how far.

One prompt gives you a band. A hundred give you a number.

Twenty answers can't pin a rate down to the point, so Citebell reports a single prompt as a band: rarely, sometimes, usually or almost always. Pool a client's prompts and the interval tightens until the figure is worth quoting, always with its interval, its sample size and how it was collected.

Mention rate

41%

95% CI 37–45% n = 1,440

120 prompts · measured via API · last 4 weeks

ChatGPT
Prompts and their bands
Prompt Named
best payroll software for a 20-person company usually · n = 12
payroll app that files taxes automatically sometimes · n = 12
payroll software with a mobile app for staff almost always · n = 12
cheapest payroll for paying contractors rarely · n = 12

and 116 more prompts

Up 6 points this week. Is it real?

AI answers drift from week to week on their own. Citebell calls a change only when the interval of the difference excludes zero. Otherwise the report says “no significant change”, so nobody presents noise to a client as progress.

This week vs last week

+6 pts No detectable change

95% CI -2 to +14 pts

Last 4 weeks vs the 4 before

+5 pts Significant

95% CI +1 to +9 pts

Then check what the answer says about your client.

A mention doesn't help if the answer gets the price wrong. Citebell checks each claim against the fact sheet the brand keeps, and traces a wrong one to the answer and the source it came from.

Answer 7 of 20 ChatGPT

Illustrative data
For a 20-person company, Northwind Payroll is a common pick: plans start at $4 per employee a month1 and it files federal and state payroll taxes automatically2. Contoso Pay is cheaper if most of your team are contractors.
1

Fact sheet: $6 per employee a month, since March 2026

Source cited: payroll-reviews.example/northwind-2024

Incorrect
2
Fact sheet: Federal and state filing included on every plan
Correct

Fix the cause. Then re-run the prompts.

Each finding is an issue with a known fix or an opportunity with an uncertain payoff, ranked by opportunity × confidence ÷ effort and shown with its evidence. When you mark one done, Citebell re-runs the affected prompts on each engine, on that engine's schedule, and reports what moved. No traffic forecasts.

Issue Show the prices on Northwind's pricing page as text, not script

Answers quote a 2024 review for the price because the pricing page draws its prices with JavaScript, which most AI crawlers don't run.

evidence: strong effort: S affects 4 prompts

  1. Perplexity first read on day 5, final read on day 35
  2. Google AI Overviews first read on day 7, final read on day 35
  3. Google AI Mode first read on day 7, final read on day 35
  4. Claude first read on day 10, final read on day 35
  5. ChatGPT first read on day 14, final read on day 35
  • Pending
  • First read: early signal
  • Final read: verified, or no detectable change

What works today, and what's coming

Collection from the engines is switching on one engine at a time, starting with Claude. Set up your clients now; the measurements above appear as each engine goes live.

Today

  • A workspace for your agency, with each client's brand, competitors and aliases
  • A fact sheet per brand: prices, features, locations
  • Prompts by topic, with CSV import, suggestions and version history
  • Your team and your clients invited with their own roles
  • Your agency's name and logo, and a log of who changed what

With collection

  • Sampled answers from ChatGPT Claude Perplexity Google AI Overviews and AI Mode
  • Bands, pooled rates with intervals, and change calls
  • Claims checked against the fact sheet
  • Ranked fixes and their verification
  • Reports to share with clients

Priced by tracked prompts

A tracked prompt is one question you measure for a client, sampled several times a week on every engine. Plans are sized by how many prompts you track across all your clients.

Prices aren't published yet. The trial tracks up to 20 prompts for 14 days.

Show clients numbers you can defend.

Measurements start as each engine's collection goes live.

Citebell · AI search visibility for agencies · Methodology

Northwind Payroll and Contoso Pay are fictional. Every figure on this page is illustrative. ChatGPT, Claude, Perplexity and Google are trademarks of their owners; Citebell isn't affiliated with them. Photographs: Bonnie Kittle and Joachim Schnürle on Unsplash.