How to Check If ChatGPT Recommends Your Business (A 20-Minute Test)

Open ChatGPT, ask it the question your best customer would ask before they'd ever heard of you, and see whether you appear. Repeat it five times, across five engines, and write down what you get. That's the test. It takes about twenty minutes and most companies have never run it.

Step 1: Write your prompts like a buyer, not like a marketer

The most common mistake is testing your own brand name. "What is [Your Company]?" tells you almost nothing useful — someone asking that already knows you exist. The prompts that matter are the ones asked by people who don't know you yet. Write ten prompts, spread across four intents:

Write these in the words your customers actually use. If your buyers say "AI agent" and you say "agentic workflow orchestration," test theirs.

Step 2: Run each prompt at least five times

This is the step everyone skips, and skipping it makes the whole exercise worthless. AI answers are not stable. Research from AirOps found that only around 30% of brands stay visible from one answer to the next, and just 20% remain visible across five consecutive runs of the same prompt. A single run tells you nothing — run each prompt five times, in fresh sessions with memory and personalization turned off, and record how many times out of five you appeared. That fraction — not a yes/no — is your actual visibility.

Step 3: Test across engines, not just ChatGPT

Different engines retrieve differently. Run the same set through ChatGPT, Perplexity, Google AI Overviews and AI Mode, Gemini, and Claude. You'll usually find you're strong in one and absent from another. The reason is almost always a difference in what each system indexes and trusts — not a difference in your marketing. While you're testing, note which sources the engines cite. If Perplexity keeps citing a particular industry directory or a Reddit thread, those are places your competitors are winning that you can enter this month.

Step 4: Record four things per run

Keep it in a spreadsheet — one row per prompt, per engine, per run:

Mentioned and cited are different things, and the gap between them is diagnostic. If you're mentioned but never cited, models know you exist but aren't pulling from your site — usually a content structure or crawlability problem. If you're cited but rarely mentioned, your content is useful but your brand isn't attached to it strongly enough.

Step 5: Run the same test for three competitors

Your own numbers mean nothing in isolation. Appearing in 2 of 5 runs is a crisis if a competitor appears in 5 of 5, and a decent starting point if nobody appears consistently. Map where your competitors are being cited — review platforms, directories, community forums — because that list tells you exactly where to invest the next 30 days.

Reading your results

Zero of 5 runs, across all engines: Start with access, not content. Check whether GPTBot, ClaudeBot, PerplexityBot and Google-Extended are blocked in your robots.txt, and whether your site renders server-side. A site that builds its content entirely in the browser is often invisible to crawlers that don't execute JavaScript — this single issue accounts for a surprising share of complete invisibility.

Appearing inconsistently, 1 to 3 of 5: You're in the consideration set but weakly. This is usually an entity and corroboration problem. Claims that appear only on your own site read as marketing rather than fact to a model. The fix is getting the same facts stated on third-party sources — directories, review platforms, press, community discussions.

Described inaccurately: Someone else may own your name in the model's entity graph, or your own site never states plainly what you do in one extractable sentence. Both are fixable and both take about 60 days.

What to fix first

In order of fastest movement:

None of these are tricks. Google has said plainly that ordinary SEO fundamentals still apply to AI search and that there's no special schema required for AI Overviews. Be sceptical of anyone selling a secret markup hack. The job is discipline and measurement, not a shortcut.