Every Hint Co. audit follows the same protocol. Same sample size, same engines, same buckets, same extraction rules.

Thirty questions, three samples per engine, three engines. 270 responses per audit.

Each audit runs against thirty buyer-shaped questions drawn from real prospect language in the category. Every question is sampled three times per engine to account for response variability, then re-run across three independent engines. Total sample size per audit: 30 x 3 x 3 = 270 responses.

Occasionally a single engine drops a response (rate limit, timeout, refused answer). The audit reports actual successful count, not the target. For example, EX-02 landed at 269 of 270.

Three answer engines, chosen for coverage.

The API tier is directionally representative of consumer chat behavior, not identical to it. All three engines expose the current web-connected variant most often used in real buyer research. Results are labeled by engine so consumers of the audit can weight them by which surface they care about.

Buyer questions are grouped by intent.

Two-stage extraction: word-boundary candidate scan, then LLM adjudication for ambiguous vendors.

Stage 1 scans each response for occurrences of every known-brand token, case-insensitive, with word-boundary anchors so “pipedrive” matches “Pipedrive” but not “pipedrivesomething.” Every hit is captured with plus/minus 200 characters of surrounding context.

Stage 2 applies only to vendors whose canonical name overlaps with a common English word (Segment, Make, Refine, Default, Outreach, Heap, Iterable, June, and similar). Every candidate hit for an ambiguous vendor is passed to Claude Haiku 4.5, which reads the context and answers a single question: is this the company, or generic usage? Only confirmed hits count toward the leaderboard. Every adjudication is logged to a cache keyed on canonical brand plus context hash, so re-runs of the reporter never re-bill.

The known-brand list is category-specific and built before the audit runs. Adding a brand mid-audit is not permitted. A brand named zero times in the audit is still worth recording as absent.

Six codes, one per confirmed brand mention.

Every mention that survives the two-stage matcher runs through a third pass. Claude Haiku 4.5 reads the response, focuses on how that specific brand is treated, and classifies the treatment into exactly one of six codes.

Pick share of total = (SOLE + PRIMARY) ÷ total responses. This is the metric that maps to revenue. A dashboard that only tracks appearance treats a "Salesforce is bloated" mention identically to "Salesforce is the recommended CRM." The classifier separates them.

Recommendation-when-mentioned = (SOLE + PRIMARY + ALTERNATIVE) ÷ total appearances of that brand. A high number here with low appearance signals a retrieval problem. A low number with high appearance signals a positioning problem. Different diagnoses, different work.

URL frequency, split by first-party vs external.

For engines that surface citations (Perplexity always, Claude with web_search often, GPT-4o Search Preview sometimes), the audit records the cited URL. Domains are then rolled up and split into two groups: the brand’s own domain (its documentation, blog, community) and external sources (independent publishers, listing sites, review sites, forums).

This split matters because it teaches which channel of content is actually doing the work. A brand that gets cited overwhelmingly from its own domain has a first-party-content playbook the answer engines respect. A brand that gets cited overwhelmingly from external listing sites has a placement problem, not a content problem.

Three explicit non-goals.

Any brand named in a published audit can request a free counter-audit.

The counter-audit runs the same protocol, on the same date, with the same engines. If it produces different numbers, both audits publish alongside. The original and the reply. The delta is annotated.

The intent is disclosure over adversarialism. The audit is a snapshot, not a verdict on any brand’s worth. A brand that wants to add context, refute a caveat, or run the same protocol with a different bucket set is welcome to.

This methodology document is versioned. Every audit references the methodology version it ran under. Material changes to the protocol bump the version number.