How Citebell measures

AI answers change almost every time you ask, so single answers and exact ranks are noise. Citebell asks each tracked prompt many times on every engine, pools the answers into rates, and reports every rate with its 95% interval and the number of answers behind it. This page is generated from the same code that computes the figures.

Version

Current methodology: M1.0

Every figure is stamped with the version it was computed with. A change to a formula, default, weight or inclusion rule starts a new version, and charts show a break where it starts. Clarifications that change no figure keep the version and are listed in the change log below.

Engines, models and sampling

Each engine is its own series, collected as its label says. Its model is pinned: a model change is a reviewed release, and charts break where it happens. Answers come through official APIs and logged-out Google results, so they aren't personalised.

Engine Model Runs per prompt per week
ChatGPT search OpenAI API + web search gpt-5.6-terra 3
Claude Claude API + web search claude-haiku-5-5 3
Perplexity Perplexity Agent API (Sonar) + web search perplexity/sonar 3
Google AI Overviews Google AI Overviews (logged-out, non-personalised) google-ai-overviews 7
Google AI Mode Google AI Mode (logged-out, non-personalised) google-ai-mode 7

Per plan: a range spans the plans' rates.

Which answers count

  • The answer was collected and stored.
  • On API engines, web search actually ran. Answers from model memory are left out and reported as a data-quality figure (search rate).
  • Headline figures use unbranded prompts only. Branded prompts feed branded visibility, a sanity check that never enters a headline.
  • Sponsored placements are never organic: they are reported apart as sponsored presence.
  • A Google results page without an AI Overview counts as not mentioned; the AI Overview presence rate is reported beside it.
  • A brand is mentioned when the answer names it (any alias), confirmed by the extractor, and counts once per answer.

Formulas

Each prompt i has a weight wᵢ = topic weight × prompt weight (1–5, default 1). p̂ᵢ is the share of a prompt's included answers that mention the brand. Figures are per engine; cross-engine figures are the equal-weight mean of the engines with a value.

Mention rate
answers mentioning the brand ÷ included answers
Visibility Score
VS = 100 × Σᵢ wᵢ p̂ᵢ ÷ Σᵢ wᵢ A weighted mean mention rate, with no hidden position or sentiment terms. Overall: equal weights over the brand's engines. Topic, intent and persona visibility are the same over their prompts.
Share of voice
SoV_b = R_b ÷ Σₖ Rₖ Rₖ = answers mentioning tracked brand k (the brand and its competitors); answers naming no tracked brand are left out.
First-mention rate
answers where the brand is the first tracked brand named ÷ answers naming any tracked brand
Top-3 rate
answers where the brand is among the first three brands named ÷ included answers
Prompt coverage
share of prompts with p̂ᵢ ≥ θ, θ = 20%
Prominence
Σ_(s names b) |s|·e^(−pos(s)/|S|) ÷ Σ_s |s| Over the sentences s of an answer (a brand-adapted position-adjusted word count), averaged over answers, 0 where the brand isn't named.
Citation rate
answers citing at least one of the brand's URLs ÷ included answers
Citation share
the brand's citations ÷ all citations, by registrable domain
Source gap list
domains cited in answers that name a competitor but not the brand, ranked by the sum of those answers' prompt weights; top 25
Net sentiment
(positive − negative brand mentions) ÷ brand mentions
Accuracy rate
answers whose checkable brand claims are all correct ÷ answers naming the brand with at least one checkable claim
Attribute association
share of brand mentions carrying each attribute phrase cluster; the 20 most common per period
Competitor co-mention
P(k | b) The share of answers naming the brand that also name competitor k.
Completeness
collected answers ÷ scheduled answers

Uncertainty

One prompt on one engine in one period uses the 95% Wilson interval, with z = 1.96. Its figure is shown as a band while the interval's half-width is over 15 points:

  • Rarely · < 20%
  • Sometimes · 20%–50%
  • Usually · 50%–80%
  • Almost always · ≥ 80%

Pooled figures (topics, the brand, the Visibility Score, share of voice) use a cluster bootstrap: resample prompts with replacement, then answers within each chosen prompt, and recompute the figure 2,000 times (seeded, so it is reproducible). The interval is the 51th and 1,950th replicate in order. This respects the correlation between runs of the same prompt, which naive pooling ignores.

A topic, intent or persona gets an interval only with at least 10 prompts with answers in the period; smaller slices show a band.

A period with less than 80% of its scheduled answers collected is marked incomplete and left out of change tests. Figures for a period settle 24 hours after it ends.

Intervals cover sampling error only. Extraction accuracy (precision and recall per engine and language) will be published here with each extractor version once the labelled evaluation set is complete.

Change detection

A change between periods is significant when the 95% cluster-bootstrap interval of the difference, paired by prompt, excludes zero; otherwise it reads "no detectable change". Only complete, settled periods in the same series are compared.

Changes listed per topic, intent or persona are filtered with Benjamini–Hochberg at a false discovery rate of 10%, one family per kind of slice. Prompt-level changes appear only as band changes.

Scores

The Visibility Score is shown with its interval, n, the engines and the methodology version, next to the per-engine scores. Share of voice, citation rate, first-mention rate, sentiment and accuracy are shown beside it, never blended into it.

AI Readiness Score

Readiness measures whether engines can access and understand a site. It does not predict visibility. It is deterministic, so it has no interval:

Score = 100 × (1 − Σ weights of failed checks ÷ Σ weights of applicable checks)

A warning counts as half a failure. A check is applicable when it passed, warned or failed; unknown, not applicable and not-checked checks are in neither sum. The score is rounded. Any failed critical check caps the score at 49.

Weights

  • Critical 10
  • High 4
  • Low 1
  • Informational 0

Bands

  • Blocked 0–49
  • Needs work 50–79
  • Ready 80–100
Check Tier Weight Evidence
RC-01 AI bots in robots.txt Critical 10 Strong evidence
RC-02 AI bots at the CDN or firewall Critical 10 Strong evidence
RC-03 Key pages can be indexed and quoted Critical 10 Strong evidence
RC-04 Content without JavaScript High 4 Strong evidence
RC-05 Sitemap Low 1 Moderate evidence
RC-06 Bing Webmaster Tools and IndexNow High 4 Moderate evidence
RC-07 Training-bot policy Informational 0 Strong evidence
RC-08 Organization schema Low 1 Weak evidence
RC-09 Structured data Low 1 Weak evidence
RC-10 Update dates Low 1 Moderate evidence
RC-11 Easy-to-quote content Informational 0 Moderate evidence
RC-12 llms.txt Informational 0 Strong evidence
RC-13 Server response time Low 1 Anecdotal evidence
RC-14 HTTPS and redirects High 4 Strong evidence

Verifying a fix

When a recommendation is marked done, its affected prompts are measured again on each engine, alongside untouched prompts of the same topics where there are any. Engines pick up changes at different speeds, so each engine has a first read after a delay that depends on the kind of fix, and a final read. Each read collects for 7 days: 5 extra runs per prompt on API engines, one a day on Google.

Engine On-site content New page Off-site work
ChatGPT search day 14 day 21 day 28
Claude day 10 day 10 day 28
Perplexity day 5 day 5 day 28
Google AI Overviews day 7 day 7 day 28
Google AI Mode day 7 day 7 day 28

Final read: On-site content day 35 New page day 35 Off-site work day 45

Only answers where the engine searched count, since answers from model memory change only with a new model. The result compares the reads with the four complete weeks before the fix: improved or worse when the 95% bootstrap interval of the change (against the control prompts when there are any) excludes zero, otherwise no detectable change. Results are pending until the first read and an early signal until the final read. Readiness fixes are verified by the next audit's check: resolved or still failing. The team can change the delays for a fix.

Claude's search reportedly draws on Brave, which has no submission tool, so brand-new pages may take longer to reach Claude, or never reach it.

Series breaks

A new series starts, and charts draw a visible break, when any of these change: the competitor set, an engine's model, a prompt's text, the brand's market, the engines enabled, or the methodology version. Change tests never span a break, and history is never silently recomputed.

Known limitations

  • Answers come through official APIs and logged-out Google results, so they aren't personalised and may differ from what a signed-in user sees.
  • Prompts are single-turn.
  • Sponsored placements are separated only where the data marks them; ads inside the ChatGPT and Copilot apps aren't returned by their APIs.
  • Intervals exclude extraction error.

Change log

  1. M1.0 ·

    Clarifications as verification is built (version unchanged, no figure changes): read length and evaluation, the pooled statistic per fix, the difference in differences against same-topic controls, how engines' outcomes combine, no timed verification for brand-wide gaps.

  2. M1.0 ·

    Clarification as the readiness score is built (version unchanged): unknown, not applicable and not-checked checks are outside both sums, only a failure caps, the score is rounded, no applicable check means no score.

  3. M1.0 ·

    Additions (version unchanged, no existing figure changes): visibility, prompt coverage and share of voice per intent and per persona; the 10-prompt interval rule applies to every slice; one Benjamini–Hochberg family per kind of slice.

  4. M1.0 ·

    Clarifications as the aggregates are built (version unchanged): first-mention counts competitors only, top-3 counts every brand named; cross-surface figures are the equal-weight mean of the surfaces with a value; an answer not yet analysed is in no figure; a change test covers only prompts whose series both periods have.

  5. M1.0 ·

    Clarifications before any figure is published (version unchanged): rank percentiles, undefined resamples dropped, the paired difference and its p-value, worked examples.

  6. M1.0 ·

    Initial draft.