Prompts. Diagnoses. Reruns. Deltas.

A versioned panel of buyer-style questions, a standardized four-assistant API panel, and retained repeat observations.

Every prompt is a real question a buyer might ask before hiring — "best NYC divorce lawyer for a spouse with a business", "how do custody rulings across state lines work", "which cosmetic surgeon in NYC does rhinoplasty for men". Atlas sends the frozen prompt panel to a standardized four-assistant panel — <code class="text-primary-300">openai/gpt-4o</code>, <code class="text-primary-300">anthropic/claude-sonnet-4.6</code>, <code class="text-primary-300">google/gemini-2.5-flash</code>, and <code class="text-primary-300">perplexity/sonar</code> — via provider APIs. This is not identical to what a given consumer sees in the ChatGPT / Claude / Gemini / Perplexity apps: consumer surfaces layer routing, retrieval, personalization, and app-side orchestration on top of the underlying models. Atlas measures the standardized API panel so results are reproducible; consumer-surface variance is a separate open question. Every completed response, exact prompt, model, and repeat denominator is preserved and published with each snapshot.

For every prompt: how often were you named? Divide by attempts. That is your citation rate.

Aggregate across the prompt set for your vertical, and you have your share of AI recommendations — expressed with a 95% Wilson confidence interval so you can tell which changes between snapshots are real and which are sampling noise. A rank of #12 with a wide CI is not the same signal as #12 with a tight one; we show you both.

Every prompt where a peer was cited and you weren't is a specific, addressable gap.

We cross-reference what the assistants name you for against what they name competitors for. For each losing prompt we surface the specific competitor page that was cited and the editorial patterns those pages share — FAQ shape, schema markup, embedded quotes, specific numbers, statute references. We do not claim these patterns caused the citation; we observe them and let you decide whether the shape is worth adopting on your own pages.

Your team publishes editorial changes against the specific gaps we surface.

For a firm, this usually means: long-form FAQ pages targeting the specific buyer questions we found you missing on, schema markup review on your existing service pages, a "questions and answers" section on your homepage. What we hand you is an editorial diagnosis — the specific gaps, the competitor pages that filled them, and the shape of the fix. What we do not do is claim that publishing any single paragraph causally produces any single citation. The retrieval side of these systems is opaque; the honest posture is diagnosis plus measurement, not "paste this and get cited".

Same prompts, same panel, on a defined cadence. The delta is what we measure — not what we claim caused it.

We hold the measurement panel and prompt set constant, so movement between runs primarily reflects changes in the underlying model outputs and retrieval behavior. That is not the same as a controlled experiment on your website — assistant routing, index freshness, provider-side model changes, and background retrieval updates can also move the numbers. We report the delta transparently, flag when it exceeds the confidence-interval width, and refuse to attribute any single citation change to any single content change unless the design supports it. If you want a controlled experiment, we can design one; that is a separate engagement.