How the AI visibility audit is measured.

The GEO score is calculated across six dimensions summing to 100 points: buyer AI visibility (/24), Google Business reputation (/20), content citability (/18), structured data (/14), multi-platform presence (/12), and AI crawler access (/12). Buyer AI visibility carries the highest weight because it is the measured outcome this report exists to establish: whether your company actually surfaces when your buyers ask AI engines their questions. The other five dimensions measure the foundations that influence that outcome.

Every report produces two scores: Current Score (measured at the report date) and Potential Score (the score the company would reach if all deterministic recommendations were implemented). Recommendations are classified as deterministic (measurable, predictable impact on the grid, e.g. robots.txt unblocking, fixing JSON-LD blocks, writing answer paragraphs) or estimated (directional impact that depends on execution, e.g. publishing expertise content, earning media placements). Only deterministic recommendations contribute to the potential score; estimated recommendations are listed without points to avoid overpromising on outcomes that depend on execution quality, ecosystem behavior, and time horizons we cannot fix.

Buyer AI visibility is measured by a battery of 24 natural buying-intent scenarios, grounded in web search and distributed across five families: comparable benchmark, service or use case, implicit need, explicit audience, and decision mode. Wording is derived from the resolved sector, geography, offer, and real buyer language; naturalness, variety, and role-alignment checks block generic or artificial batteries before execution. The battery runs against three AI engine families (Anthropic's Claude, OpenAI's ChatGPT, and Google's Gemini), in their production configuration at generation time, without personalization or history. Identified competitors run through the same battery for comparative positioning: we measure "among the questions your buyers ask, who surfaces", not "who surfaces in each competitor's own queries." Agency-supplied questions are probed in a separate, clearly labeled lane and never enter the score, because questions written by an interested party steer the result.

Each scenario is executed three times per engine to absorb generation variability. 16 scenarios measure provider discovery: the company must be named or recommended. 8 scenarios measure expertise visibility: the company must be named or its site cited. A scenario is confirmed only when a majority of eligible engines produces the expected result; an outcome on one engine is retained as a partial signal, never a confirmed win. Probes that hit transient errors are retried up to two additional times before being dropped; when an engine becomes unavailable after retries, it is excluded from the measurement and the disclosure names it.

Content citability (/18) evaluates whether the site's visible content can be extracted and quoted by an AI engine. It combines visible-content depth (word-count tiers), Article markup hygiene, and the text's extractable structure: self-contained answer paragraphs of 80 to 180 words that read as complete sentences, distinct subheadings phrased as buyer questions, and an entity-definition sentence opening the homepage. FAQPage and HowTo markup are observed but score no points: they are retired rich-result formats, and visible content is what carries citation. Only pages actually accessible to the scan are evaluated: substitute content from a third-party directory is never scored, and an inaccessible site is reported as "not assessed" rather than penalized on measurements that never happened. The exact excerpts that earned points are kept in the report as an audit trail.

Structured-data validation (/14) verifies that each JSON-LD block parses, contains an @type, and has the required fields for that type (e.g. name and url for Organization, headline and author for Article). Invalid blocks do not contribute to the score regardless of their count. A complete Organization block with sameAs (LinkedIn link) and address earns a bonus.

Google Business reputation (/20) is graded: listing verification, review volume in tiers, average rating, then comparison against the best named competitor; a listing clearly out-reviewed by a competitor loses points and triggers the corresponding recommendation. Multi-platform presence (/12) counts platforms detected on the site plus the verified Google listing, also relative to the named competitors.

AI crawler access (/12) parses robots.txt (the GPTBot and ClaudeBot training crawlers, the Google-Extended control token, and the search-time retrieval agents) and verifies server rendering of the homepage (a site whose content only exists after JavaScript execution is unreadable to most AI crawlers and sees this dimension capped). Google-Extended changes neither Google Search indexing nor ranking, nor presence in AI Overviews, but it controls use of content for Gemini training and certain Gemini answers grounded with Google Search. The presence and quality of llms.txt are retained as an unscored interoperability observation and never trigger an automatic recommendation. This dimension also acts as a gate: when two or more major retrieval agents are blocked, the structured-data and citability scores are capped, because content a retrieval agent cannot read cannot be cited, regardless of its quality.

Sources cited by AI engines in their responses are extracted by an automated parser and categorized by domain (company site, LinkedIn, press, directories, official Quebec professional-order registries, social platforms, competitor sites, other). This distribution drives the recommendations. A platform showing zero citations does not mean the company has no presence there: it means the engines did not surface it for the probed questions.

Limitations disclosed: (1) Grid tiers are calibrated on a Quebec SMB cohort; upper tiers remain to be validated against higher-profile accounts. (2) Perplexity is excluded from the probe: its generation-retrieval coupling does not measure the same construct as the three probed families. (3) LinkedIn is never accessed directly (Terms of Service and Loi 25 compliance); LinkedIn presence is inferred via the site's sameAs markup and via URLs cited by the engines. (4) Content-structure analysis covers the homepage and a bounded set of inner pages; citable content outside that perimeter is not detected. (5) Estimated-impact recommendations do not contribute to the potential score. (6) The scoring grid is versioned: scores are only comparable within the same grid version. Because the page corpora analyzed for earlier reports were not retained, those reports do not contain the new passage-structure signals required by v3.1. The first complete v3.1 report is therefore the new baseline for future comparisons; no cross-version recomputation is presented as equivalent.