Tracking AI recommendations means running the same questions on the same platforms on a schedule and recording how often your business is named, because a single answer is unreliable. The bottom line: track the share of runs that name you, per platform and per question, against named competitors, monthly, with enough runs that a change means something, and use either a manual sheet, an official API with web search, or a packaged tool.
See also: answer engine optimization, explained step by step.
What to track
- Share of runs that name you. Per platform, per question, and overall.
- Share of runs that cite your site. A citation is a link to your domain.
- Competitor shares. The same numbers for your two or three closest competitors.
- Sources cited. Which domains appear when competitors are named.
- Sentiment or framing. Whether you are named as a top choice or a mention.
How many runs make a change real
A share is a proportion, and small samples swing. With 100 runs and a true share near 30 percent, the 95 percent margin of error is about 9 percentage points, so a move from 30 to 36 is within noise. With 20 runs the margin is about 20 points. SparkToro and Gumshoe ran each prompt 60 to 100 times per platform to see stable patterns. Decide in advance what change you will act on, and gather enough runs to detect it.
Three ways to run the tracking
| Approach | How it works | Best for | Limit |
|---|---|---|---|
| Manual sheet | You run questions in the apps and record results | Small question sets and first baselines | Slow, and easy to skip |
| Official APIs | OpenAI's Responses API with its web search tool, Perplexity's Sonar models, and Gemini with Google Search grounding return answers and sources for a script to analyze | Scheduled, repeatable runs at scale | API answers can differ from what people see in the apps, so treat them as a proxy |
| Packaged tools | Ahrefs Brand Radar, Semrush, PromptRush, and similar services run large prompt sets | Ongoing monitoring without building anything | You depend on their prompt sets and refresh rates |
Steps to set up tracking
- Choose 20 to 50 real customer questions. With place and situation.
- Choose the platforms. ChatGPT, Gemini, Perplexity, and Google AI features at minimum.
- Decide the number of runs. At least five per question per platform monthly.
- Create the record. One row per run with question, platform, date, named, cited, position, and other names.
- Run on a fixed schedule. The same week each month.
- Chart shares over time. Show your share and competitors' shares.
- Review the cited sources. They tell you where to build presence.
- Verify with the apps. If you use APIs, spot check against the consumer apps.
What to do with changes
If your share rises after you publish or fix something, note the change and the date. If it falls, check crawler access and page changes before assuming a platform shift, and remember that platforms change too: Semrush documented a sharp drop in ChatGPT's Reddit and Wikipedia citation share between August and September 2025.
Questions people ask
How do I track whether AI recommends my business?
Run the same customer questions on several platforms on a schedule, record whether you are named and cited in each run, and track your share of runs over time against competitors.
How many runs do I need?
Enough that a change is larger than the noise. With 100 runs and a share near 30 percent the margin is about 9 points. SparkToro and Gumshoe used 60 to 100 runs per prompt per platform.
Can I automate AI visibility tracking?
Yes. OpenAI, Perplexity, and Gemini offer APIs with web search, and packaged tools run large prompt sets. API answers may differ from the consumer apps, so spot check them.
Why did my AI visibility drop?
Check crawler access and page changes first. Platforms also shift: Semrush documented a sharp drop in ChatGPT's Reddit and Wikipedia citation share in 2025.