Test how AI models see your brand.

No catch. No email, no account, no paywall. Haide runs this same five-prompt test on every prospect call before we pitch strategy. Publishing it costs us nothing and makes the conversation sharper when someone does end up hiring us. Use it whether or not you ever contact Haide.

Parametric = the model answers from training data only, with web search switched off. It reveals whether the brand is baked into the model's world model. Dynamic = the model searches the live web, grounded in current sources. Running both and comparing is the point of the kit. The gap between them tells you whether you're winning on historical training data, current web discourse, both, or neither.

ChatGPT, then Claude, then Perplexity. ChatGPT has the largest training footprint and the most visible consumer surface, so its answers tend to match what a buyer would actually see. Claude is the best tie-breaker — its training cutoff and reasoning style differ enough to surface blind spots. Perplexity is web-grounded by default, so it mostly answers the dynamic question.

Monthly. Parametric answers only move when a model is retrained or fine-tuned, which happens every few months for the major labs. Dynamic answers move as often as your web footprint does — faster for brands running active PR, slower for everyone else. Monthly gives you enough signal to catch drift without burning time on noise.

No. The free tier on all three will run these prompts. The web-search toggle you need for the dynamic variant is available on every tier, sometimes under a different label (Claude calls it 'search the web', ChatGPT shows a globe icon). If the free tier rate-limits you mid-scan, wait or switch models.

That's a result, not an error. It means you have zero parametric presence on that model — the brand is not in training data. Run the dynamic variant next. If dynamic also returns nothing useful, the brand has neither historical training coverage nor enough current web footprint to surface. That's the gap the GEO Checklist closes.

Yes. The prompts don't verify ownership. Run them for a prospect, a competitor, a portfolio company — any brand where you want a quick read on how the three major AI models position them. Good pre-call research before a strategy meeting.

Only for prompts four and five. The first three work with just a brand name. For prompts four and five, you need a category (e.g. 'project management SaaS', not 'software') and a use case (e.g. 'remote team standup', not 'productivity') for the model to have enough context to judge authority or recommendation fit. Generic categories produce noisy answers.

The first five prompts test what the model already believes about your brand. The bonus card tests what the model extracts when it actively searches your site and reads two specific pages. The JSON blocks inside that prompt are ChatGPT's own internal tool-call syntax — GPT-5 parses them and runs a real site: query, then opens the first two results. Watch which pages rank first on that site: search and how ChatGPT summarises them. The pages that surface are the ones doing your AI-visibility work. The ones that don't, aren't. Prompt structure surfaced by Lily Ray, originally shared by Olivier de Segonzac (RESONEO) and Chris Long (Nectiv).

The JSON format used in that prompt is OpenAI's internal tool-call syntax — specifically search_query and open with turn0search ref_ids. GPT-5 parses it and executes real tool calls. Claude and Perplexity have different tool architectures, so they treat the JSON as plain text and will not trigger the same search-and-read behaviour. Perplexity always searches by default anyway, so the trick is less useful there.