Build a fixed set of about twenty prompts a real customer would type, run them every month across ChatGPT, Gemini, Perplexity and Google AI Overviews in a logged-out session, and score each answer on four levels: not mentioned, mentioned, recommended, cited with a link. Log the results in one sheet. The number that matters is the trend over three to six months, not any single answer — these systems are non-deterministic, so one good response proves nothing.
We have written about how to get your business mentioned by ChatGPT, Gemini and Perplexity. The obvious next question is the one almost nobody answers: how do you know whether any of it worked? There is no Search Console for AI assistants, so this is the method we run instead — it is manual, it is cheap, and it is repeatable.
Why can't you just check once?
Because the same question asked twice can produce different answers. AI assistants are non-deterministic, they personalise on your history, and they vary by region and by which model version you happen to get. A single check tells you almost nothing.
Three rules make the results comparable month to month:
- Always run logged out, or in a temporary chat, so your own history does not flatter the result.
- Never change the prompt wording. The moment you rephrase, you have lost the comparison.
- Run the whole set on the same day each month. Consistency beats frequency.
How do you build the prompt set?
Twenty prompts, written the way a customer types, not the way a marketer does. Split them across four groups.
| Group | How many | Example shape |
|---|---|---|
| Category + city | 6 | "best digital marketing agency in Rajkot" |
| Problem-led | 6 | "my Facebook ads are not getting leads, who can help in Gujarat" |
| Comparison | 4 | "agency or freelancer for a small business in India" |
| Direct brand | 4 | "what is Safar Spectrum Media" |
The direct brand prompts matter more than they look. If an assistant describes your business wrongly — wrong city, wrong services, confused with another company — that is a factual problem to fix before worrying about rankings at all.
What scale should you score on?
Four levels. Anything more granular becomes a judgement call and stops being comparable.
- 0 — Not mentioned. You are invisible for that query.
- 1 — Mentioned. Your name appears somewhere in a list.
- 2 — Recommended. The assistant actively suggests you, with a reason.
- 3 — Cited. It links to your page as the source of a claim.
Citation is the outcome worth optimising for, because it does two jobs at once: it sends a visitor, and it establishes that your page rather than a directory is the underlying authority. That is the whole point of GEO — publishing things specific enough that a summary has to attribute them instead of absorbing them.
Also log who else appeared. The competitor list an assistant produces is the cleanest market map you will get for free, and it is usually not the list you expected.
What should the tracking sheet contain?
One row per prompt per engine per month. Six columns is enough:
- Date · Engine · Prompt · Score (0–3) · Competitors named · Notes.
Twenty prompts across four engines is eighty rows a month and takes about an hour. Then track two numbers only: your average score, and your appearance rate — the share of prompts where you scored 1 or higher. Those two lines over six months tell you everything a dashboard would.
What else corroborates it?
Manual testing tells you what the assistant says. Three other signals tell you whether it is producing anything.
- Referral traffic from AI domains in GA4 — chatgpt.com, perplexity.ai and similar. Volumes are small; the trend is the signal.
- AI crawler hits in server logs. If assistant crawlers are not fetching your pages at all, nothing downstream can work.
- The oldest signal of all — ask new enquiries how they found you. "ChatGPT suggested you" is now a real answer, and it is the only one that confirms revenue rather than visibility.
There are paid AI-visibility trackers that automate the prompt runs. They are useful at agency scale. For a single business, a spreadsheet and an hour a month gives you the same trend line and a much better understanding of why it moved.
How long before anything changes?
Three to six months before a trend is readable, and longer for competitive categories. Assistants draw on crawled pages, third-party sources like directories and review sites, and training data that updates on its own schedule — so a single new page moves nothing.
What does move over a quarter or two is unglamorous: consistent business information everywhere, specific first-hand content worth quoting, and presence in the third-party sources these tools already trust. For a local business like Twing Cafe, the questions worth tracking are not abstract — "healthy cafe in Rajkot" is a query a real customer types, and whether an assistant names the cafe is a measurable, fixable fact. The same logic applies across the 10 industries we work in, and it sits alongside the traffic shift we covered in AI Overviews took my traffic — what now.
If someone promises AI visibility results in weeks, they are selling something. Run the sheet, be patient, and look at the trend at month six — that is the whole discipline, and it is also our approach to SEO and GEO work generally.
Key Takeaways
- One check proves nothing — AI answers vary by session, history, region and model version.
- Use a fixed set of ~20 customer-worded prompts across category, problem, comparison and brand questions.
- Always run logged out, never change the wording, run on the same day each month.
- Score 0–3: not mentioned, mentioned, recommended, cited with a link. Citation is the outcome worth chasing.
- Track two numbers only: average score and appearance rate. Log which competitors appear — it is a free market map.
- Corroborate with AI referral traffic, crawler hits in server logs, and asking new enquiries how they found you.
- Expect three to six months before the trend is readable. Nobody moves this in weeks.
Before You Ask
How do you check whether ChatGPT recommends your business?
Build a fixed set of about twenty prompts a real customer would type — category plus city, problem-led questions, comparison questions and a direct brand question — then run the same set every month across ChatGPT, Gemini, Perplexity and Google AI Overviews. Run them in a logged-out or temporary session so your own history does not bias the answer, and score each result on four levels: not mentioned, mentioned, recommended, or cited with a link. The number that matters is the trend across months, not any single answer.
What does it mean to be cited by an AI assistant?
There are four distinct outcomes and they are worth separating. Not mentioned means you are invisible for that query. Mentioned means your name appears in a list. Recommended means the assistant actively suggests you with a reason. Cited means it links to your page as the source of a claim. Citation is the strongest outcome because it survives the answer — it sends a visitor and it signals that your page, not a directory, is the underlying authority for that information.
How long does it take to change AI search visibility?
Realistically three to six months before a trend is readable, and sometimes longer. AI assistants draw on a mix of crawled pages, third-party sources such as directories and review sites, and training data that updates on its own schedule. Publishing one good page does not move it. What does move it, over a quarter or two, is consistent business information everywhere, specific first-hand content worth quoting, and presence in the third-party sources the assistants already trust. Anyone promising results in weeks is selling something.
*This is the internal method Safar Spectrum Media uses to track AI assistant visibility for clients. Timelines are our own observed ranges, not a guarantee — AI assistants change their sources and behaviour frequently. The 41+ brands, 10 industries and 25+ ad accounts figures are SSM's own, as of September 2026.