AI search competitor analysis: A repeatable benchmark

Traditional competitor research asks who ranks, earns links, or captures clicks. AI answers add another layer: a competitor may be recommended without a link, your page may be cited without your brand being recommended, or an independent publisher may supply the evidence for both brands. A defensible benchmark preserves those differences instead of compressing them into one visibility percentage.

This workflow extends an AI visibility audit into a controlled brand comparison. It uses stable buyer prompts, full-answer evidence, explicit counting rules, and a decision ledger so the next content or technical task can be traced to an observed answer.

1. Define the decision before choosing competitors

Start with one buyer task, such as choosing reporting software for a small agency. Do not mix unrelated categories, markets, or funnel stages in one benchmark.

Create a competitor set with three roles:

• Direct alternatives: products a buyer could select instead of yours.
• Answer competitors: brands that repeatedly appear for the chosen questions, even if sales does not treat them as direct rivals.
• Source competitors: publishers or product pages that answers cite to support a recommendation.

Keep these roles separate. Wikipedia, Reddit, an industry publication, and a software vendor can all appear in the same answer, but they do not compete for the same outcome. Limit the initial comparison to your brand plus three or four meaningful product competitors. A giant list makes qualitative review shallow and changes the denominator whenever a new name appears.

Record approved spelling variants, product names, and domains for each brand. Do not treat a common word as a brand match without checking the answer context.

2. Freeze a buyer-prompt panel

Use questions a buyer would actually ask, not prompts designed to force every brand into the answer. A compact panel can include:

1. category discovery: “What tools help a small agency monitor visibility in AI answers?”
2. task fit: “Which platform is best for comparing competitor mentions across AI search?”
3. constraints: “What is a practical option for a three-person marketing team?”
4. comparison: “Compare Brand A and Brand B for multilingual monitoring.”
5. evidence: “Which tools show the sources behind AI answers?”

Separate unbranded discovery prompts from comparison prompts that name brands. Direct-brand prompts can test accuracy, but including them in discovery share of voice will inflate the named brand by design.

For every observation, preserve the exact prompt, platform or surface, language, target market, collection time, and full answer. If you change the prompt wording, platform, or market, start a new comparison cohort rather than silently extending the old one.

3. Use a four-layer scorecard

Count only successfully collected answers, and show the denominator for every rate.

Report results as auditable statements: “Brand A was recommended in 6 of 20 successful answers” is interpretable. “Brand A has 30% AI visibility” is not, unless the report defines the prompt set, platforms, weighting, failures, and counting rule.

Do not merge these layers. A brand can lead mentions but trail recommendations. A competitor domain can receive citations because its documentation answers a technical question, while another product receives the recommendation. That is a content finding, not a contradictory result.

4. Compare answers in AEO Mantis

In AEO Mantis, use your monitored buyer prompts as the fixed panel. Review the exact answers behind the summary states, then compare which configured competitors appear, how they are framed, and which domains or pages support the answer. Keep excluded or direct-brand prompts available for qualitative review, but do not put them into an overall visibility rate unless your counting rule explicitly calls for them.

For each prompt, record:

• your brand’s mention, recommendation, and citation state;
• every configured competitor that appears;
• the first meaningful description of each brand;
• cited owned pages and recurring third-party sources;
• collection failures or answers with no visible sources.

This is where product data ends and editorial judgment begins. The tool can preserve answers, mentions, citations, competitors, and sources; a reviewer still has to decide whether a description is accurate, whether a cited passage supports it, and what action is justified.

5. Inspect the source gap, not just the brand gap

For an important answer, open every visible source and identify what it contributes. Classify the page as owned documentation, product marketing, independent editorial, community discussion, or reference material. Then check whether the cited passage supports the nearby claim.

A useful source-gap record contains:

• the prompt and answer reference;
• cited URL, final destination, title, publisher, and check date;
• the claim the source appears to support;
• whether support is clear, partial, contradictory, or unverified;
• the closest page on your site;
• a proposed action and owner.

The presence of a source does not reveal a private ranking formula. Treat it as evidence from one observed answer. Repeated source patterns across the fixed panel deserve investigation; a single citation is a lead, not a mandate to copy the page.

6. Turn gaps into bounded actions

Assign each verified gap to one of four queues:

• Accuracy: correct an outdated or incomplete description on an owned page, or prepare a factual correction for a third-party publisher.
• Coverage: improve the existing URL when it already serves the buyer intent; create a new page only for a genuinely distinct task.
• Evidence: add current documentation, methodology, examples, or limitations that you can substantiate.
• Access: investigate crawlability, canonical, robots, indexing, or visible-page problems before rewriting content.

Give every action an owner, source observation, expected output, and next review date. “Improve AI visibility” is not a task. “Add the missing export-limit documentation to the existing migration page and review the same 20-prompt cohort next month” is.

Avoid assuming that schema, llms.txt, or allowing a training crawler guarantees inclusion. Search and citation retrieval controls differ by platform, and eligibility does not guarantee selection. Verify the relevant surface, crawler, and index state separately.

7. Work through an illustrative example

Illustrative example only — these numbers are invented, not AEO Mantis or customer results. A team tests 20 successful English answers across two AI search surfaces using ten fixed buyer prompts. Its brand is mentioned in 7 answers and recommended in 3. Competitor A is mentioned in 11 and recommended in 8. Competitor B is mentioned in 6 and recommended in 4.

The raw gap favors Competitor A, but source review makes it actionable. Six of its eight recommendations cite a current implementation guide. The team’s closest page describes the product but omits setup time, data prerequisites, and a limitation buyers repeatedly ask about. The decision is to update that existing page, not publish a synonym article.

The team records both rates with denominators, preserves the six answers and source passages, publishes the verified update, and reruns the same cohort after a defined interval. A later change would be an observation, not proof that the edit caused it; models, source indexes, and answer behavior may also change.

8. Rerun without moving the goalposts

Compare equivalent completed periods. Keep prompt wording, platforms, market, language, brand aliases, competitor set, and failure handling stable. If the business requires a new competitor or prompt, version the cohort and report the break in comparability.

Track publication, crawling, indexing, mention, recommendation, citation, referral visit, and conversion as separate states. An IndexNow response can confirm receipt of a URL submission; it does not prove indexing. A citation can occur without a click, and a visit cannot automatically be attributed to a particular AI answer.

The purpose of AI search competitor analysis is not to manufacture a leaderboard. It is to find a small number of defensible differences—who appears, how they are framed, what evidence supports them, and what your team can improve—then measure again under the same rules.