Blog · Competitive monitoring in AI search
Competitive AI Monitoring: Track When You Lose Recommendations
Priya Bothra · September 12, 2026
Competitive AI monitoring tracks how often AI assistants recommend you versus competitors for the same buyer questions, and alerts you when that balance shifts beyond normal variation. The key is separating real losses from noise: AI answers vary between runs, so you need baselines, rolling averages and thresholds before you react. When a real loss appears, the cited sources usually reveal why.
Losing a recommendation is rarely dramatic. A competitor publishes a strong comparison page, earns coverage in a publication models cite, updates its review profiles or ships a feature that answers a common constraint. Over a few weeks, AI answers start favoring them. Without monitoring, you notice months later in pipeline data. This guide explains how to set up monitoring that catches shifts early and how to respond.
What to monitor
| Metric | Why it matters for competition |
|---|---|
| Recommendation share by competitor | Shows who AI assistants favor for each prompt cluster |
| Visibility rate by competitor | Shows who is in the consideration set |
| Cited sources per cluster | Shows which pages and domains shape answers |
| Reasons given for recommendations | Shows which attributes the models associate with each brand |
| New brands appearing | Shows emerging competitors early |
Track these by model and prompt cluster. Aggregated totals hide the most useful signals.
Separating signal from noise
AI answers are probabilistic. SparkToro and Gumshoe found fewer than 1 in 100 runs from ChatGPT or Google's AI returned the same brand list. Week-to-week swings of several points are normal, so build noise tolerance into monitoring.
Build a baseline
Collect four to six weeks of repeated sampling before setting alerts. Calculate the average and typical variation for each competitor, cluster and model.
Use rolling averages
Compare a rolling two- or four-week average with the baseline instead of reacting to a single week.
Set thresholds based on sample size
With a few hundred answers per cluster per period, a change of about 10 percentage points is a reasonable alert threshold. With fewer answers, set wider thresholds or pool clusters. A practical rule: alert when the rolling average moves outside the range you observed during the baseline period, and the move persists for two consecutive periods.
Use a control
Track a set of stable prompts where nothing competitive has changed. If those move too, a model update is the more likely cause.
Common causes of lost recommendations
| Cause | How it shows up | Where to look |
|---|---|---|
| Competitor published new content | Competitor URLs start appearing in citations | Cited sources for the cluster |
| Competitor earned third-party coverage | New publication or review page cited | New domains in citations |
| Review profile changes | Review site pages cited with new ratings or descriptions | Review platforms |
| Your content became outdated | Old facts in answers, fewer citations of your pages | Accuracy log, your cited URLs |
| AI crawler access problem on your site | Your pages disappear from citations across clusters | Server logs, robots.txt, CDN rules |
| Model or retrieval update | Many brands shift at once, control prompts move | Control set, provider announcements |
| Market change | A new constraint, such as a regulation, appears in answers | Reasons given in answers |
Public research explains why source changes matter so much. Different assistants rely on different sources: Profound's citation analysis found Wikipedia led ChatGPT citations while Reddit led for Perplexity and Google AI Overviews. A competitor gaining presence in the source your category depends on can move answers quickly in one model and not another.
Root-cause analysis in five steps
- Confirm the loss is real. Check rolling averages, sample sizes and the control set.
- Isolate where it happened. Which model, which cluster, which competitor gained?
- Compare cited sources before and after. New URLs or domains usually explain the change.
- Read the reasons. How do answers now describe the competitor and you? Look for new attributes, such as "better for enterprise" or "cheaper."
- Check your own side. Look for outdated pages, crawler errors or changes you made.
The response playbook
| Root cause | Response |
|---|---|
| Competitor's new comparison content | Update your comparison page with current, specific facts and honest trade-offs |
| New third-party coverage favoring competitor | Offer the publisher current information about your product; pursue coverage in comparable sources |
| Review profile shift | Encourage genuine reviews from satisfied customers; respond to criticism; update profile facts |
| Your content outdated | Refresh affected pages and third-party profiles |
| Access problem | Fix crawler rules and bot protection immediately |
| Model update | Re-baseline; avoid overreacting to shifts that affect everyone |
| Genuine product gap | Share the insight with product teams; state honestly which buyers you fit |
A note on reviews: the FTC's rule on fake reviews prohibits fake reviews, including AI-generated ones, and paying for reviews conditioned on sentiment. Responses must stay within those rules.
Setting up alerts and routines
- Alerts: notify the program lead when recommendation share for a priority cluster crosses the threshold for two periods.
- Weekly: review alerts and new brands appearing in answers.
- Monthly: review competitive share trends by model and cluster, and cited source changes.
- Quarterly: full competitive audit and competitor set review.
Common mistakes
Reacting to single answers. One screenshot of a competitor recommendation is not a trend.
Monitoring only one model. Losses often start in one assistant.
Tracking mentions, not recommendations. Being listed while a competitor is recommended is still a loss.
Ignoring new entrants. AI assistants sometimes surface emerging brands before they appear in your sales conversations.
Copying competitor tactics blindly. Understand why they won before responding.
A hypothetical example
A hypothetical CRM vendor's recommendation share for "best CRM for real estate agents" in Perplexity falls from a baseline average of 34% to 21%, holding for three weeks. Control prompts are stable, so a model update is unlikely. Cited sources show two new Reddit threads and a recently updated comparison article, both highlighting a competitor's new mobile app. ChatGPT results are unchanged. The vendor updates its own comparison page with its mobile capabilities, provides current information to the comparison publisher, and asks its customer community team to share accurate information where appropriate. It tracks the cluster weekly to see whether the share recovers.
How Bob Builds AI helps
Bob Builds AI's Visibility Monitoring compares recommendation share, citations and positioning against competitors across AI models and tracks recommendation changes over time. Its Slack integration can route alerts to the right channel.
FAQ
How do I know if a competitor is winning AI recommendations?
Track recommendation share for you and competitors across a stable set of buyer prompts, sampled repeatedly across the AI assistants your buyers use. Compare rolling averages with a baseline and look for sustained shifts rather than single-answer changes.
How often should I check competitor AI visibility?
Review alerts weekly, trends monthly and run a full competitive audit quarterly. Increase frequency around major launches, pricing changes, competitor announcements and AI model updates.
Why did my brand stop being recommended by ChatGPT?
Common causes include new competitor content, new third-party coverage favoring competitors, outdated information about your brand, crawler access problems on your site and model or retrieval updates. Comparing cited sources before and after the change usually reveals the cause.
What is a normal fluctuation in AI visibility?
It depends on sample size and category. With repeated sampling, week-to-week swings of several percentage points are common. Establish your own baseline over four to six weeks and alert on changes that exceed that range for more than one period.
Can AI model updates change which brands are recommended?
Yes. Model and retrieval updates can shift answers across many brands at once. Tracking a control set of prompts helps distinguish model updates from competitor-driven changes.
How should I respond when a competitor overtakes me in AI answers?
First confirm the shift is real and find its cause through cited sources and answer reasons. Then respond at the source: update your content, correct outdated information, pursue comparable third-party coverage and state clearly which buyers you fit best.
Should I track new competitors that AI assistants mention?
Yes. AI assistants sometimes surface emerging or adjacent brands before they appear in sales conversations. Adding them to your monitoring set early helps you understand how the market is changing.
Conclusion
Competitive AI monitoring is about catching real shifts early without chasing noise. Baselines, rolling averages, thresholds and a control set tell you when a loss is real, and cited sources tell you why. Most losses trace back to a specific source change that you can respond to.
Start by baselining recommendation share for your top three prompt clusters against your top three competitors for a month. Once you know what normal looks like, alerts become meaningful. Bob Builds AI can automate the monitoring and alerting across models.