← Leaderboard

Methodology

helloranked measures how visible brands are in the answers of major AI assistants. The methodology is public by design; numbers you can't interrogate are numbers you shouldn't trust.

How we measure

Each industry has a fixed, versioned panel of ~50 questions phrased the way real people ask AI assistants ("What CRM should a fast-growing startup use?"), each with paraphrase variants. Panels are immutable once frozen: any change creates a new version and is annotated on trend charts. Every measurement cycle (currently monthly), each prompt variant is sent to each measured model 3 times (LLM answers are non-deterministic; single samples produce noise).

Measured models

From answers to scores

Each answer is analyzed by a two-stage extraction pipeline (alias matching, then an LLM judge that confirms brand identity, disambiguates common words like "monday" or "On", and records rank, sentiment, and whether the brand was explicitly recommended). Per brand and model we compute:

• Mention rate — share of answers mentioning the brand (weight 45%)
• Position score — how early the brand appears when mentioned, mean of 1/rank (weight 25%)
• Recommendation rate — share of answers explicitly recommending it (weight 20%)
• Citation rate — share of search-mode answers citing the brand's own site (weight 10%; ships with search-enabled measurement, currently 0)

The per-model score is the weighted blend, 0–100. The Global Visibility Score weights each model's score by its assistant's approximate consumer market share:

Market-share weights (v2)

Statistical confidence

LLM answers are non-deterministic samples, so every score carries sampling error. We quantify it with a percentile bootstrap: each model's responses are resampled with replacement 1,000 times, the score recomputed each time, and the middle 95% of outcomes reported as the confidence interval (the ± you see next to scores, and the shaded band on trend charts). Run-over-run movement gets a solid arrow only when the bootstrap confidence interval of the change excludes zero. Smaller changes appear as muted hollow arrows — the direction is shown for transparency, but the movement was indistinguishable from sampling noise and should not be read as a trend.

Share of voice & industry traffic weights

The global leaderboard ranks by share of voice: within each industry, a brand's share of the category's total AI mentions (its slice of the conversation relative to competitors), multiplied by that industry's share of consumer advice-seeking demand, then expressed as a share of the whole index. This rewards dominating your category over merely being present in a big, crowded one; within-industry leaderboards remain unweighted. Mention share is measured on the market-share-weighted mention rate (a mention on a more-used assistant counts for more). The demand weight: since no AI vendor publishes query distributions, demand is proxied by real search behavior — for each industry we take a small, consistent set of advice-seeking queries built by prefixing the category's name and synonyms with “best”/“top” (“best vpn”, “best vpn service”, “best credit card”…), and sum each query's 12-month average monthly US search volume. Using the identical “best <category>” template for every industry is deliberate: it measures the specific recommendation question — the closest proxy to what people ask an assistant — rather than the size of a category's broader keyword universe (which over-weights heavily-monetized verticals like credit cards). Averaging over a full year keeps seasonal categories (tax software peaks in spring) comparable. Because even advice-seeking demand is heavily skewed — one category can be 40%+ of the total — the per-industry demand is square-root compressed before weighting: this damps the spread so no single high-demand category dominates the ranking, while preserving the order of industries by demand. The assumption (advisory search correlates with assistant-question demand) is documented, and weights are versioned and source-cited like the model weights. An unweighted by-industry-score view stays one click away.

Search-enabled modes & citations

Perplexity answers are always web-search-backed, and for a pilot set of industries we additionally measure ChatGPT and Gemini with their web-search tools enabled. For these answers we record the pages the assistant cited. A brand's citation rate — the share of search-backed answers citing the brand's own domain — feeds the 10% citation component of that mode's score. The "+ web search" variants are reported separately and are not part of the Global Visibility Score (their market-share split from the base assistants isn't measurable yet); Perplexity, being inherently search-backed, remains in the global blend. Run-over-run movement remains mention-based. Cited-domain panels on brand and industry pages come from this data.

Known limitations (roadmap)

• Search-mode measurement covers a 3-industry pilot; remaining industries are raw-model only for now.
• Cross-industry score comparisons are weaker than within-industry ones: each industry has its own prompt panel.
• Suite products are assigned to one primary industry (e.g. HubSpot lives in CRM, not Email Marketing).
• Market-share weights above are marked with their sources; placeholder entries are replaced as sourced data lands.

FAQ

Questions about the methodology? It's versioned in the open — every raw answer is stored and re-analyzable.