How to track your brand's AI search visibility (and why one check isn't enough)

⏰ The 30-second version

• Checking your AI visibility once tells you almost nothing. Model updates move results, and so does the content you publish.
• Track three numbers: mention rate, model spread, and question coverage. Everything else is decoration.
• Monthly is enough for most brands. Weekly if you're actively publishing.
• Keep the question set fixed. If you change the questions, you're measuring the questions, not your progress.
• The point isn't a dashboard. It's knowing which question to write about next.

Why one check expires

A visibility check is a reading, not a fact. Three things move it:

1. Models update. Training data and behaviour change on a schedule you don't control.
2. Competitors publish. The slot you almost held gets taken by whoever answered the question better.
3. You publish. The only variable you own — and the only way to know it worked is to have measured before.

That last one matters most. Without a baseline, you can't tell whether the article you spent a week on did anything.

The three numbers worth tracking

1. Mention rate — how often you're named

Out of your fixed question set, in how many answers does your brand appear? Fifteen questions, four mentions, mention rate 27%. That's your headline number.

Track it per model, not blended. A blended average hides the situation that actually kills brands: strong in one model, invisible in another.

2. Model spread — how many AIs know you

Count how many of ChatGPT, Claude and Gemini name you at all.

• 3 of 3: durable. Rare.
• 2 of 3: healthy, with a known gap to close.
• 1 of 3: fragile. Your visibility depends on which app the customer opened.
• 0 of 3: invisible, and it's a content problem rather than a product one.

When we measured 73 well-known beauty brands, only one held a top position across all three models. Most brands that appeared at all appeared in one.

3. Question coverage — which needs you own

Mention rate says how much. Coverage says what for. List the specific questions where you appear:

Question	You	Who else appears
sunscreen for sensitive skin	✓	2 competitors
vegan skincare brand	✗	3 competitors
moisturizer for dehydrated skin	✗	4 competitors

This table is the actual output of tracking. The rows where you're absent and competitors aren't — that's your content queue, already prioritised.

A tracking routine that survives contact with a busy week

Set it up once. Write 10–15 questions the way customers phrase them, with no brand names. Save them. This list should not change; a fixed set is what makes months comparable.

Run the same pass each time. Ask every question in each of the three models. Record: appeared or not, and which competitors did. Fifteen questions across three models is about 45 prompts — roughly an hour by hand.

Monthly is the default cadence. Move to weekly only while you're publishing actively and want to see whether it landed.

Review with two questions:

• Did the mention rate move after what I published?
• Which un-owned question has the most competitors in it? (That's the one worth writing next — competitors clustering there means the demand is real.)

What not to bother with

• Chasing a single answer. Ask the same question twice and phrasing shifts. Rates over a set of questions are stable; individual answers aren't.
• Screenshot collections. They feel like evidence and can't be compared month to month.
• Vanity totals. "Mentioned 40 times" means nothing without the denominator.
• Optimising for one model. You'll win it and lose the other two, and your customers use all three.

The honest version of the effort

An hour a month, done by hand, gets you all three numbers. That's genuinely fine, and worth doing before you buy anything.

If you'd rather not run 45 prompts by hand, that's the chore Bulti removes — enter a brand name and get the questions you're losing across models, with what to publish first. Manual or tool, the habit is the same: fixed questions, several models, regular cadence, and a decision at the end of each pass.

Honest footnotes

• Figures referenced here come from our July 2026 measurement of 73 well-known Korean beauty brands using a fixed 15-question set across ChatGPT (gpt-5), Claude and Gemini.
• This approach measures recommendations from trained model knowledge. Live-retrieval tools such as Perplexity behave differently and are worth tracking separately if your customers use them.
• "Appeared" means the brand name was present in the answer text. Borderline cases exist.
• Results move as models update; treat every reading as a snapshot.