The direct answer: ask the engines your buyers' questions, score the answers, and repeat weekly. That's the whole method. You can run it by hand in an afternoon with a spreadsheet, or automate it. This guide covers both — including the part most vendors skip: why a single pass can lie to you.
1 — Write down the questions your buyers actually ask
Not your keywords — questions. "Best invoicing software for a small remote team?", "QuickBooks alternative for freelancers?". Ten is enough to start; 30 covers a category. Two rules from our measured studies: phrase them the way a buyer types, and freeze the set. If you change questions every week, you can't tell whether the answer moved or you did.
2 — Run them on five engines, in clean sessions
ChatGPT, Google AI Overviews, Gemini, Claude, Perplexity. Fresh chat each time — history contaminates answers. The same brand, same week, can be the default on one engine and invisible on another: in our payments study one brand measured 93% recommendation share on one engine and 53% on another. One engine is never a proxy for the rest.
3 — Score every answer on four things
| Record | The question it answers |
|---|---|
| Named | Does your brand appear at all? |
| Recommended | Named approvingly — or just mentioned in passing? |
| Rank | First pick, or fourth of four? |
| Citations | Which sources did the engine read to build this answer? |
The fourth row is the one people skip, and it's the actionable half: the cited sources are where the answer comes from, which makes them the levers. Tally the hosts across all your answers. The sources cited repeatedly that never name you are your citation gaps — usually the fastest fixes available.
4 — Repeat the run before you trust it
Here's the part that invalidates most screenshots: AI answers vary run to run. Across our published studies, roughly 15% of identical question × engine checks changed verdict on re-ask. And in our retrieval study, ChatGPT showed no sources at all on half of the Chinese-language checks. Run every check at least twice, score the consensus, and treat any single-pass result — good or bad — as noise until confirmed.
5 — Re-measure weekly, or the data rots
Sources update, competitors publish, engines re-index. A baseline is a photo of a moving object. A weekly cadence on a frozen question set is the minimum that turns "what does AI say about us" into a metric you can act on: baseline, action, re-measure, prove the flip.
When to automate, and with what
If you only want a snapshot, do it by hand — it's free, and you'll learn your category's sources on the way. If you want the weekly loop, use a tool: we keep an honestly-scored comparison of AI visibility trackers, including where each one is weak. Ainswer is ours: fixed question sets, five engines, multi-run consistency, and the citation tally built in — one question is free, no email.
Where these numbers come from
Figures cited are from Ainswer's published multi-run studies: SaaS payments (235 scored answers), retrieval & local brands (307), and B2B supplier shortlists (120). Raw verdicts are included in each report.
Skip the spreadsheet. See your number.
Paste your site — one real buyer question across five engines, free, no email. The full scan shows your share, your engines, and the exact sources to fix.