How we score AI search readiness
The Citevera headline overall score is a weighted rollup of a five-stage AI decision funnel. Each stage scores how the audit items in that stage map onto the question the stage is asking. The headline is the funnel; the funnel bars on your report are the inputs.
The five stages
Listed in funnel order, with the rollup weight applied to the headline. Weights chosen via N=50 production scan calibration (2026-04-26) and revised on 2026-10-19 against 283 scans (Trust 10% to 15%, Coverage 20% to 15%; scans scored before that date keep the earlier split). See the notes below.
Detection
Can AI find and identify your site at all? Robots.txt access for GPTBot, ClaudeBot, PerplexityBot, and Gemini. Presence of llms.txt and llms-full.txt. Sitemap reachability. Crawl-time response codes. The first stage of the funnel because nothing downstream matters if AI cannot reach the page.
Understanding
Once AI fetches the page, can it parse what is there? Heading hierarchy, schema.org coverage (Organization, Person, FAQPage, HowTo, Article, Product), meta tags, content extractability. The audit prompt asks specifically whether the structure makes the topic explicit, not whether the page exists.
Coverage
Does the site demonstrate enough breadth on the topic that AI engines treat it as authoritative on the subject? Number of pages per topic cluster, internal-link structure, hub-and-spoke organization, content depth on the primary topic. The 33x rule comes from Sites with 50+ blog posts averaged 1,373.7 AI crawler visits vs 41.6 for sites with no blog - a 33x gap. (Duda, February 2026, 858k sites on a single platform; single-platform, correlational) Methodology and sources: citevera.com/ai-search-statistics.
Conversion
When an AI engine refers a visitor to the site, can that visitor act? Working CTAs, visible pricing or contact paths, schema-marked Products or Services with Offers, no broken forms. Conversion is downstream of detection and understanding - if AI never sends the visitor, the conversion path does not matter, but if it does, this is where the funnel either closes or leaks.
Trust
Does the site carry enough trust signal that AI engines treat it as a credible source rather than just a relevant one? sameAs linkage to LinkedIn, Crunchbase, Wikidata; Organization schema; Review, Rating, and Testimonial markup; verifiable identity. Since the 2026-10-19 Trust revision (formula funnel-v2) the stage also reads author and expert attribution, sourced statistics, freshness signals, about and contact pages and case studies, a Google Business Profile only counts as a strong signal for businesses with a physical location, and a stage measured by only one or two findings is pulled toward the middle instead of scored to an extreme. The same day its headline weight went from 10% to 15%, taken from Coverage, now that the stage tracks the rest of the funnel (correlation 0.62 to 0.77, up from 0.38 to 0.50).
Why your headline number changed
On 2026-04-28 we replaced the old headline formula with the funnel rollup described above. The old formula was a weighted average of three holistic axis ratings (40% AEO readiness, 40% GEO readiness, 20% crawlability) the audit model emitted alongside the audit items. We discovered those axis ratings ran systematically optimistic compared to the audit items themselves - across a calibration sample of 50 production scans, 76% of headlines moved by more than 5 points and 64% moved by more than 20 points when re-derived from the audit items via the funnel rollup.
The audit items are the ground truth - they are the concrete present, weak, and missing signals the audit surfaces, and they are what we ship as fixes. Trusting the holistic axis ratings over the audit items meant the headline was disagreeing with itself. The new formula reads the audit items directly so the headline reflects what the audit actually says.
Historical scans keep their stored headline so trend charts before the cutover are unchanged. Scans run on or after 2026-04-28 use the new formula. Trend charts render a small dotted boundary where the formula transitions, so a March-vs-May comparison does not silently mix two formulas. The audit items, fix recommendations, and per-stage funnel bars are unchanged - only the headline math moved.
Re-tuning cadence
We re-audit the weights when any of the following holds: the production scan population exceeds ~200 complete scans on the current formula, the audit prompt changes meaningfully (new categories, revised severity rubric), or six months elapse since the last calibration. The Trust stage was re-weighted on 2026-10-19 against 283 production scans (formula label funnel-v2; scans scored before that date keep their funnel-v1 numbers, and the trend chart marks the boundary). The same revision raised the headline weight of Trust from 10% to 15% and lowered Coverage from 20% to 15%: four candidate splits were swept over the 283 scans and all landed within 3 headline points of each other, so the choice was structural - keep Detection and Understanding on top and weight the three downstream stages equally. The re-tune playbook lives in the repository at scripts/calibration/README.md.
Per-stage status bands
Each stage carries an emotional status label alongside its 0-100 number, derived from the band the score falls into. Same bands apply to the headline overall score.
- 86 to 100: dominant - the strongest possible reading on this stage.
- 71 to 85: competitive - this stage is working in your favor.
- 51 to 70: understandable - present but with room to push higher.
- 21 to 50: weak - real gaps are surfacing here.
- 0 to 20: not usable - the audit items in this stage describe a critical absence.
Questions or pushback? Email the team. The audit and the methodology are auditable; we want the disagreements.