How we measure, and what we don’t claim.

1. Questions, not keywords

Each site tracks 15 to 60 buyer questions, depending on the plan: the questions someone asks an AI assistant before choosing a provider like you. We suggest them from your site and competitors; you edit them.

Each question has its own language and country. Questions never contain your brand name: a question only your brand can answer tells you nothing about whether you’re recommended.

2. Repeated samples on every engine

Every week we ask each question three times on each engine in your plan: ChatGPT, Gemini, Google AI Overviews, Google AI Mode and Perplexity on every plan, plus Claude on Business (Claude once a week: each Claude answer runs a web search, which makes it the most expensive query). Grok is a paid add-on, asked once a week per question in your main country only. Answers are collected through a data provider’s APIs (DataForSEO) and, for Grok, xAI’s own API, not by logging into consumer apps.

Where each answer comes from: ChatGPT and Gemini are read as a user sees them in the app, with its searches and sources. AI Overviews and AI Mode come from Google’s results page. Perplexity, Claude and Grok are asked through each model’s API with web search turned on, not through perplexity.ai, claude.ai or grok.com: the answer and its sources are close to what the app shows, but not identical.

Why three: public research by SparkToro and Gumshoe found less than a 1 in 100 chance that ChatGPT or Google’s AI returns the same list of brands twice. A single answer is anecdote.

SparkToro: AIs are highly inconsistent when recommending brands

3. Mentioned, cited, and a range

• Mentioned: your brand name or an alias appears in the answer text.
• Cited: the answer links to a page on your site as a source.
• Rates are the share of answers that mention or cite you, shown with a 90% Wilson confidence interval. With few answers the interval is wide, and we show it wide.
• A change is only called “up” or “down” when it’s larger than the noise. Otherwise we say “no clear change yet”.
• We never report a “position in ChatGPT”: answers vary too much for a single position to mean anything.
• For each answer we also store the competitors mentioned and every cited URL, grouped into your pages, competitor pages and third-party pages.

4. Measuring the effect of what we do

The cleanest signal is direct: a page we published starts being cited for its question. We show that with the date it was first cited.

Rate changes after an action are shown next to the action, with their intervals, but we don’t claim they were caused by it. Engines change, competitors publish and seasons move.

5. How articles are written and verified

• Each article answers one question with a measured gap or real demand in Search Console.
• A research step gathers sources from the web; claims in the article link to them.
• An automatic verification step flags numbers without a source and lists where your brand ranks itself first. Articles with any error stay as drafts.
• Your brand voice comes from a style guide learned from 3–5 of your own pages, which you can edit.
• The cover image uses your brand colours and contains no text. In the body, images only when they inform: a chart of figures the article cites (with the source written under it), a diagram of its steps or a photo from your own site. An AI illustration only for an abstract idea, never products, people or real places. Internal links come from your real sitemap. Each article has a table of contents and FAQ schema.

6. Automatic publishing, with safeguards

• Articles that pass verification with zero errors are scheduled and published automatically after an approval window (3 days by default). You get an email with a one-click “Don’t publish”.
• The first 3 articles on each site are approved by hand, to calibrate tone and quality.
• At most one automatic publication per site per day: a steady cadence, no bursts.
• Any article can be unpublished in one click for 7 days. You can pause a site or switch to drafts only at any time.

7. Our link policy

We don’t exchange links, buy links, run a network of sites or place links automatically. Google’s spam policies list link exchanges, paid links and automated link creation as link spam, and we won’t put your domain at that risk.

What we do instead: find the third-party pages that AI engines cite for your questions and that mention your competitors, draft honest outreach for you to send, track its status, and suggest legitimate directories for your sector. You decide and you send.

Google Search spam policies: link spam

8. The data we store

• Measurements: the questions, the answers, the cited URLs and the costs of collecting them.
• AI traffic script (optional, under 1 KB): sends the page URL and the referrer. No cookies, no IP address stored, no visitor ID. We keep daily counts per page and per AI source only.
• AI crawler hits are only visible for sites served through Lantis, our sister product, as daily counts per bot.
• Free report: the domain, the email you give us and a salted hash of your IP to limit abuse.

9. Limits we admit

• We measure fresh, non-personalised answers. Logged-in users with memory, location or history may see something different.
• Visits from Google AI Overviews and AI Mode arrive as Google and can’t be separated. Apps that don’t send a referrer are lost, so AI traffic is a lower bound.
• Three samples a week is enough to see real shifts, not small ones. Narrow intervals need more weeks.
• Most AI citations point to third-party sites. Publishing on your own site is only part of the work.
• We haven’t published a results study yet. When we have enough data we’ll publish one with intervals and a comparison group, not before.