What we measure, what we change, and what we don’t claim.

1. Sampling you can reproduce

You choose the buyer questions. They are a measurement panel, not search volume and not real conversations. Each sample keeps the question text, channel, method (consumer UI, official API or search API), model version when available, market, time and cited links.

• API output is never labelled as what a consumer saw in a chat app.
• Missing or rate-limited samples are shown as missing and left out of the denominator. They are never counted as “not mentioned”.
• Retries create new runs. History is never overwritten.

Live audits currently sample two channels through their official APIs with web search: GPT-5 mini and Perplexity Sonar (via OpenRouter). Each audit is capped by a monthly model budget.

2. Edits grounded in your facts

Every opportunity splits what we observed, what we think explains it, and what is still uncertain. Drafts can only use facts you confirmed. A rule-based check flags numbers, platforms, integrations, compliance terms and absolute words (“unlimited”, “best”, “no watermark”) that aren’t backed by a fact or already on the page. You can’t approve until they’re resolved.

Structured data suggestions must match visible text. We never invent reviews, ratings, customer counts, prices or capabilities.

3. Approval that means something

Approval is tied to the SHA-256 hash of the exact saved text and the facts it relied on. Editing the text or changing a fact clears approval. Undo brings back the words, not the approval. Downloading the same approved version again isn’t counted as a new delivery.

4. Verification, not assumption

Exporting is not publishing. After you publish, Citelane fetches the public URL (your page only, no private networks, no getting around bot protection) and checks that every approved sentence is present. A match proves the words are live. It does not prove indexing, AI citation or revenue.

5. Results without made-up causality

At day 7, 14 and 28 you record a hypothesis, a main metric, anything else that changed at the same time, and a decision: continue, adjust or stop. Before/after differences are observations. Only a proper controlled test can support a causal claim. We show no composite “GEO score”, and anything not measured stays “not measured”, never 0.

Source: Google Search Central, “AI features and your website” (checked 30 Sep 2026). This applies to Google’s AI features only.