How CiteFuel Scores AI Visibility — Full Methodology

Why we publish this: the GEO/AEO space has a snake-oil problem. Tools publish opaque "AI scores" with no explanation of inputs, weights, or validation. We publish ours because (a) transparency builds trust, (b) you should understand what you're optimizing for, and (c) if our weights are wrong, we want to hear about it. Email support@citefuel.com with disagreements. Product scoring changes are recorded on this page. Separate, dataset-based publications live in CiteFuel Research and follow our editorial policy and corrections process.

Score 2.0 (current engine): your published GEO score is a blend of two layers — Readiness (R), the deterministic on-site check table below (26 checks worth 129 points, R = earned ÷ available × 100, same partial/skip rules as before), and Visibility (V), a live-sampled presence rate: we generate a disclosed synthetic prompt set about your category, sample an AI-answer engine across N runs, and report the share that produced a positive brand-presence reading — as a rate with a 90% bootstrap confidence interval. A text adapter generally counts a brand appearance; a supported source/reference match may also count. This is not a uniform recommendation or citation rate. Individual outputs vary, so repeated fixed-prompt sampling and an interval quantify, but do not remove, uncertainty. Your published score is a confidence-weighted blend of R and V — nominally 0.65×R + 0.35×V, but V's actual weight scales with how many samples it's based on (see "Confidence-weighted blend" below for the full formula and why). If Visibility wasn't sampled this run (rate-limited, errored, or genuinely not yet run), the score is R alone, and the report says so explicitly — never a guessed or assumed V.

Google-specific boundary: Google says its AI features use the same foundational Search requirements, require normal indexing and snippet eligibility, need no special AI schema, and ignore llms.txt. See Google's AI features guidance and generative AI optimization guide. CiteFuel's checks and scores are internal measurements, not Google ranking signals.

Hard gates: two implemented failures are severe enough that no amount of Readiness or Visibility credit offsets them: the HTTPS/TLS check fails, or an AI assistant asked cold about the brand states a claim that clearly contradicts the site's own evidence. Either condition caps the published score at 39 (grade D), regardless of what R and V would otherwise blend to. The WAF row is a CiteFuel-origin user-agent response comparison, not verified crawler-network evidence; it remains coverage-critical and can lower Readiness, but it is not a 39/D hard-cap condition.

The 6-engine visibility matrix

Visibility (V) attempts up to 6 AI-answer engines. Authentication, billing, rate limits, and provider health can change which lanes complete — this configuration snapshot is not a promise of readings:

A report's sample count only counts completed readings. An unavailable engine is neither a zero nor a pass; it is reported as unavailable and excluded from the measured denominator. The report's own engine, timestamp, locale, N, confidence interval, and provider status are authoritative for that run.

Confidence-weighted blend — why your V weight isn't always 35%

The nominal split is 65% Readiness / 35% Visibility, but Visibility's actual weight in your headline scales with how many samples it's based on — a single unlucky or lucky sampled answer can swing a pooled presence rate by a large amount. Repeated answers can vary even for a fixed prompt, so a 10-sample free-tier read shouldn't count for as much as a robust 48-sample paid-tier read.

confidence = min(1, n_runs ÷ 20) — N ≥ 20 is the full-confidence weighting threshold, where Visibility reaches its full nominal 35% weight. Below that, V's effective weight is 0.35 × confidence and R picks up the difference, so the two always sum to 100%. At 0 samples this degenerates to R alone (100% Readiness, 0% Visibility) — the same behavior as before Visibility sampling existed at all. At 20+ samples it's byte-identical to the nominal 65/35 split.

Visibility here means positive brand-presence readings, not a uniform citation or recommendation rate. Confidence-weighting prevents a small sample from taking the full nominal 35% of the formula. For example, Readiness 100 and Visibility 0 at pooled N=10 yields 0.825 × 100 + 0.175 × 0 = 82.5. That low-N output is provisional: it is not evidence of a grade or certification and is not comparable with an N ≥ 20 full-confidence audit. Reports disclose the effective weight and N; the separate evidence gate also prevents low-N weighting from producing a new S/90+ certification.

Coverage and S/90+ certification gate

The displayed formula score is not enough to earn an S grade. New reports can publish 90+ only when the unrounded formula is at least 90.0 and all of these are true: at least four visibility engines produced readings; pooled N ≥ 20 (the full-confidence weighting threshold); at least 90% of Readiness check weight was actually measured; and no critical crawl/index/citation measurement was skipped. Visibility must also be fresh: a cached reading cannot earn a new S/90+ certification. A robust badge additionally requires N ≥ 40. If display rounding produces 90 while the raw formula is below 90.0, or if an evidence gate fails, the report shows the raw unrounded value, display-rounded value, and gate reasons, then caps the published score at 89/A. Missing evidence never becomes successful evidence.

Every report lists measured weight, skipped checks, healthy engines, attempted engines, unavailable provider reasons, pooled N, confidence interval, and the exact gate reasons. Cached results disclose the original Visibility measurement time and cache age. Authority coverage is reported separately in the SEO track; an OpenPageRank not_found response means authority is unknown, not zero and not passed.

Mention rank & cited-source rank

Alongside the presence rate, each engine's reading includes two rank signals computed from the same sampled answers (no extra API calls):

• Mention rank — your brand's position among every brand name mentioned in the engine's answer, ordered by which name appears earliest in the text. Rank 1 means you were the first brand named; a higher number means competitors were named before you. No rank at all means your brand never came up in that answer.
• Cited-source rank — your domain's position within the engine's own list of cited sources/references for that answer (when the engine surfaces one), same ordering convention. This is a different signal from mention rank: an engine can name your brand without ever citing your site as a source, or vice versa.

Your report shows the median of each rank across every sampled run for that engine — a single run's rank is noisy, so the median across N runs is less sensitive to one outlier. It still carries sampling uncertainty and should be compared across identically configured audits.

Every check below still keeps the same partial/skip rules as before: a partial result earns half the check's weight, and a skipped check (e.g. AI sampling temporarily unavailable, or a Wikipedia article that doesn't exist for your brand) is excluded from the denominator. Coverage and skips are disclosed separately so a smaller denominator is never mistaken for stronger evidence.

SEO Track B — a separate, thin sibling score

Alongside the GEO score above, every audit also computes a lightweight SEO score — classic on-page and technical search-visibility signals, worth 95 points across 12 checks. It's reported separately (report.seo) and never mixed into your GEO score or grade — the two are scored independently. Checks marked "reuses your GEO result" read an existing GEO check's verdict rather than re-running it.

App AI-Visibility — a separate scale, for iOS App Store links

Paste an App Store link instead of a domain and CiteFuel runs a different, purpose-built 14-row audit registry, never mixed into the website GEO or SEO scores above. The current app_v2 headline normalizes 10 scored Readiness rows worth 75 available points, then blends that Readiness result with the separately sampled Visibility layer. A partial earns half weight and a skipped scored row is excluded from the available denominator. Two developer-site rows — optional app schema and the llms.txt/FAQ/answer-first inventory — are reported but informational and earn zero Readiness credit. The two live sampling rows, ai_recommendation_sampling and competitor_sov, describe one shared Visibility sample and also earn no fixed Readiness points. The historical 100-point flat checklist remains in report JSON only for explicitly labeled legacy comparison.