To track AI model responses about your company, define 10 to 20 buyer prompts, run them across the major engines on a fixed schedule, score every answer for mentions, top picks, sentiment and citations, then compare the numbers week over week. One-off checks are not enough. The same AI query changes its answer roughly 70% of the time (SparkToro), so a single screenshot tells you almost nothing.
There is a reason method posts dominate this query. In Honeyb's own citation dataset, the leading rival guide on this topic was cited 16 times by AI engines, and it is one of only two pages in the niche whose brand actually gets named in the answers. That is unusual: 62% of AI citations never name the brand at all (Semrush). Structured, repeatable methods get cited, and occasionally credited. The method below is the one we run ourselves.
Why single checks mislead
On 13 July 2026 we ran a controlled measurement: 20 buyer prompts, each run three times, across four engines via API (gpt-5-mini, gemini-2.5-flash, claude-haiku-4-5 and Perplexity sonar). That produced 240 answers. Between identical runs, the top-pick brand changed 44% of the time on Gemini, 43% on Perplexity, 35% on ChatGPT and 28% on Claude.
Top-pick change rate
How often the top recommendation changes between identical runs
Engines also disagree with each other. In our measurement, two engines named the same top brand on only 20% of prompt pairs at the low end (ChatGPT and Perplexity) and 53% at the high end (Gemini and Claude). Even the full set of brands mentioned only overlapped 42% to 67% between identical runs. Check one engine once and you are sampling a distribution, then calling it a fact. This is the core argument in our guide to how to monitor AI brand mentions.
The four-step method to track AI model responses about your company
### Step 1: Define 10 to 20 buyer prompts
Write the questions your buyers actually type, in category language rather than brand language: "best [category] for [use case]", "alternatives to [incumbent]", "[category] pricing". In our measurement each answer surfaced 4.8 to 5.2 brands on average, so category prompts reveal your position on the competitive shortlist, not just whether you exist. Ten prompts is the practical floor; 20 gives you enough coverage to segment by intent stage.
### Step 2: Run them on a schedule, across engines, more than once
Run every prompt on at least the engines your buyers use, on a fixed weekly cadence, and run each prompt more than once per cycle. Averaging three runs smooths the 28% to 44% run-to-run flip rate into a usable signal. Note the interface difference too: via API, all four engines returned sources on 100% of answers in our test, but consumer apps may show fewer, and ChatGPT and Gemini often hide citations in their consumer interfaces (Semrush).
### Step 3: Record four metrics per engine
| Metric | Definition | What a change means |
|---|---|---|
| Mention rate | % of prompts where your brand appears anywhere in the answer | Your baseline visibility in the category |
| Top-pick rate | % of prompts where you are the first or recommended option | Your share of the default recommendation |
| Sentiment | How the model frames you when it does mention you | Positioning risk, independent of visibility |
| Citation share | % of cited sources you control or influence | Which pages are feeding the answers |
Want to see this in action?
See how every major AI model talks about your brand. Free to start.
Mention rate and top-pick rate together are effectively your AI share of voice; we cover the calculation in how to measure AI share of voice. Citation share is the leading indicator: in our measurement ChatGPT drew on 445 distinct domains across 900 citations, and the top three domains took just 7.6% of them, so the citation pool is wide open.
Citation mix also differs sharply by engine. Perplexity leaned on community sources in our test, with Reddit at 71 of its 498 citations (14%) and YouTube at 40 (8%). That is consistent with Semrush's wider finding that Reddit is the most-cited source in AI answers at 40.1% of all citations. Track where each engine sources its answers, not just whether you appear.
### Step 4: Compute week-over-week deltas against known noise
A delta only means something if it clears the noise floor. Here is our own readout from the 13 July 2026 measurement, which doubles as a reference noise table for each engine.
| Engine | Top-pick change between identical runs | Brand-set overlap between runs | Avg sources per answer |
|---|---|---|---|
| ChatGPT (gpt-5-mini) | 35% | 42% | 15.0 |
| Gemini (2.5 Flash) | 44% | 54% | 10.7 |
| Claude (Haiku 4.5) | 28% | 67% | 8.8 |
| Perplexity (sonar) | 43% | 61% | 8.3 |
Read it like this: if your Gemini top-pick rate moves ten points in one week, that sits inside the 44% run-to-run flip rate and is probably noise. If your Claude mention rate falls for three consecutive weeks, that is a real trend, because Claude was the most stable engine in the set. Averaged runs plus a noise reference turn anecdotes into a defensible readout.
Spreadsheet or tool
The whole method runs in a spreadsheet at 10 prompts and one engine. It stops scaling at around 20 prompts, four engines and three runs, which is 240 answers a week to score by hand, exactly the size of our one-day test. Entry pricing for dedicated trackers, from each vendor's public pricing page on 13 July 2026:
| Tool | Entry price | Prompts at entry | Trial |
|---|---|---|---|
| Honeyb (our product) | $29/mo | 10 prompts scanned daily, 1 model | No trial; free AI visibility check |
| Otterly | $29/mo ($25 annual) | 15 prompts, 4 engines | Free trial |
| Sight | $49/mo | 10 prompts, 5 engines | 7-day trial |
| LLM Pulse | EUR 49/mo | 50 prompts, 5 surfaces | 14-day trial, card required |
| Dageno | $79/mo | 50 prompts, 3 platforms | Free trial |
Whichever route you take, keep the readout format identical to the manual version above, so you can audit any tool against your own spot checks. We walk through what a production readout looks like in our AI model tracking dashboard guide.
Start the baseline this week: 10 prompts, two engines, three runs each, one spreadsheet tab. You will know your mention rate and top-pick rate by Friday, and you will have a noise floor by the following week. If you want the first readout without the manual work, run a free AI visibility check and see what the engines currently say about your company.
