How to Track Competitors in Google AI Overviews and AI Search

To track competitors in Google AI Overviews and AI search, define the competitor set before collection, freeze a balanced cohort of unbranded customer questions and record every valid response using the same rules. For each brand, capture text mention, role, order where meaningful and visible source links separately. Compare matched prompts, markets and dates; never treat a failed run as an absence.

The output is competitive evidence from a controlled sample, not a universal ranking. It should answer which alternatives enter consideration, on which topics, in what context and with what supporting domains. It should not claim to reproduce every response Google or another AI engine shows to every user.

What competitor tracking covers

Competitor tracking in AI search is the repeated observation of a declared group of brands across prompts that represent a specific buying decision. The unit is one response to one prompt under recorded conditions. A competitor mention is a name or approved alias in the answer text. A competitor citation is a visible link to a domain classified as belonging to that company. These signals can co-occur, but neither implies the other.

This is not general brand monitoring. A general program can scan many themes for any reference to your company. A competitor program holds the comparison universe and decision context stable so differences are interpretable. The wider operational method belongs in the guide to tracking brand mentions across AI search.

Google's current AI features guidance, reviewed 21 July 2026, explains that AI Overviews and AI Mode may use query fan-out to explore related subtopics and sources. That is one reason a conventional rank for a single query cannot describe the generated comparison. The same guidance says ordinary technical requirements and SEO best practices remain relevant; competitor tracking is measurement, not a special route to inclusion.

Build a defensible competitor set

Begin with the buyer's choice, not the companies your leadership notices most. Define audience, country, language, problem, company size and constraints. Then form three labeled groups:

• Direct alternatives solve the same problem for the same buyer with a comparable product or service.
• Substitutes solve the need differently, such as an internal process, marketplace or adjacent category.
• Emerging entrants appear often enough in sales, customer research or the baseline sample to warrant observation.

Write an inclusion and removal rule. For example, a direct alternative may need documented availability in the market and confirmation from sales or customer research. An emerging entrant might join the main set only after recurring in two scheduled review periods; that threshold is a governance choice, not an industry benchmark. Keep a broader discovery watchlist separate so a surprise brand can be noticed without silently changing the core denominator.

Create a canonical identifier, official name, product names, spelling variants and unambiguous abbreviations for every company. Exclude generic words and ambiguous initials. Store approved domains independently from aliases. A review-site page about a competitor is third-party evidence, not an owned competitor citation.

Design prompts for fair comparison

Use unbranded prompts across problem discovery, category education, comparison and selection. Add real constraints—market, integration, regulation, scale or use case—where buyers use them. A prompt that names one competitor is diagnostic and should not enter the unbranded competitive share calculation because it prompts the very presence being measured.

Balance the cohort so a single product strength does not dominate. If one competitor specializes in enterprise deployments and another serves small teams, report those segments rather than declaring one globally more visible. Freeze the wording during a reporting window and version additions. When the cohort changes, show both the matched-prompt series and the expanded current view.

A six-step collection and classification workflow

1. Record the execution context

Save exact prompt, engine and surface, language, market, timestamp, device or access method and any search mode. If conditions cannot be matched, report them as separate segments. Do not preserve private customer or account information in the shared dataset.

2. Preserve the full response

Store the complete answer before extracting brands. Keep collection errors, refusals and empty outputs as explicit states. Screenshots help audit layout, while structured text supports repeatable classification. A timeout does not mean that no competitor appeared.

3. Detect brands using the frozen alias table

Mark presence for each canonical competitor and retain the matched passage. Review ambiguous matches rather than counting them automatically. Version any alias correction and, when material, recompute earlier periods using the same rule.

4. Assign a role, not just an order

Useful roles include primary recommendation, suitable for a stated constraint, alternative, factual example, comparison subject and warning. Record list order only when the answer truly presents an ordered list. Do not invent first, second and third positions for prose that recommends different options for different needs.

5. Record sources separately

Save visible URLs, display labels, placement and normalized domains. Classify each as focal-brand owned, competitor owned or independent third party. A cited competitor documentation page can support a passage without recommending that competitor; read the answer before assigning meaning.

6. Aggregate matched observations

Calculate mention rate as valid responses containing the competitor divided by valid responses in the same cohort. Show the denominator and failure count. Other measures can include primary-recommendation rate, co-mention rate, owned-source presence and topic coverage. The guide to AI search visibility metrics explains the formulas and reporting cautions.

Illustrative competitive example

Imagine a procurement platform tracking itself and four direct competitors across 36 unbranded prompts. It runs the cohort in two English-speaking markets, creating 72 planned observations. Four fail, leaving 68 valid answers. Competitor Cedar appears in 23, Northstar in 17 and Harbor in 12. Those counts do not yet make Cedar “the market leader”; they describe mention presence in this illustrative sample.

Role analysis shows Cedar appearing mainly in educational lists, while Northstar is recommended more often when prompts mention public-sector procurement. Harbor has fewer mentions but a high share of visible links to its own documentation. The useful actions differ: examine why the focal brand lacks public-sector evidence, review source clarity and compare topic coverage. Copying Cedar's pages would not follow from the data.

Suppose the next period adds twelve integration prompts and Northstar's overall rate rises. The report should show the original 36-prompt matched trend separately. Otherwise the team may mistake a cohort tilted toward Northstar's strength for a change in the engines.

Reading Google outcomes alongside answers

Google says traffic from its AI features is included in Search Console's overall Web performance report. Search Console can therefore help inspect impressions, clicks and landing pages on your verified property, but it is not a prompt-level competitor archive or a separate AI Overview competitor report. Do not attribute every search-performance change to AI features or compare your private Search Console figures with a competitor's unknown denominator.

Use owned-site outcomes as a second layer. A competitor may appear frequently while your relevant page still receives qualified visits; another may earn visible source links without being recommended. The response sample explains representation, while site analytics explains observed owned outcomes. Neither alone establishes causality.

Quality controls and limitations

• Require a versioned competitor and alias registry.
• Keep prompted-brand diagnostics out of discovery rates.
• Audit a sample of labels with a second reviewer.
• Separate valid non-mentions, failed runs and unavailable source areas.
• Show matched-cohort trends when prompts or markets change.
• Attach representative answers to every material movement.

No method captures every response shown to every user. Results can vary with wording, context, region, interface and time. The tracker cannot prove why a model chose a competitor, whether a cited page caused a mention or whether visibility caused a purchase. It can identify repeatable patterns and evidence gaps in a declared sample.

Use the free AI visibility audit to establish a starting view, then preserve the cohort and rules if you continue. The durable advantage is not a favorable snapshot; it is a competitive measurement system whose boundaries, denominators and evidence remain clear.