Why does AI recommend my competitors?

By Abiot Y. Derbie - Founder, RankEcho - Updated 2026-09-01

What did the answer actually show?

Begin with the narrowest factual statement: on a recorded date, a named system returned an answer to an exact prompt, and that answer presented a tracked competitor in a defined way. Preserve whether your brand also appeared, whether either brand had a visible citation, which URLs were linked, and whether the answer called the rival a recommendation, an example, an alternative, or merely mentioned it.

Do not convert that record into 'the engine prefers the competitor.' The output might change across repeats, accounts, locations, product modes, or index updates. A cited URL can inform a source hypothesis, but it does not reveal every undisclosed input or prove why the system selected the brand.

• Exact prompt, engine and surface label, date, locale, account state, and repeat number.
• Full answer text and visible cited URLs before domain normalisation.
• Your-brand state: cited, named without a link, absent, or ambiguous.
• Competitor state: recommended, compared, listed, merely mentioned, or ambiguous.
• Unavailable and error cells retained separately from completed misses.

Which competing hypotheses should you test?

Use the table as a starting set, then add product- and market-specific explanations. Evidence can support more than one hypothesis, and the correct outcome may remain unknown. A disconfirming check should be capable of weakening the hypothesis; a checklist that only collects confirming examples is not a diagnosis.

How should the Find stage capture a competitor loss?

Register the prompt panel and the replacement rule before scoring. One defensible rule is: your brand is absent and a tracked competitor is explicitly recommended or presented as a suitable option. A rival merely named in the user's prompt, a citation title, a disclaimer, or a long undifferentiated list should not automatically count.

For rates, the denominator is all eligible completed prompt-engine-repeat cells in the stated panel. Report the numerator and denominator for your brand, each rival, and each engine; keep unavailable cells outside the rate but visible in the report.

• Separate cited, mentioned, compared, listed, and recommended outcomes.
• Review ambiguous brand-name matches and similarly named companies manually.
• Map visible citations to the exact cell without claiming an undisclosed causal source.
• Retain repeat-level answers before calculating any combined rate.

What does a worked synthetic diagnosis look like?

Suppose five commercial prompts run on two engines with two scheduled repeats: 20 planned cells. Two are unavailable, leaving 18 eligible. A tracked competitor meets the registered replacement rule in 7 of 18 cells (38.9%); your brand is named in 2 of 18 (11.1%). Four of the seven replacement cells visibly cite the same third-party category page, and that page does not list your product.

That observation supports an external-coverage hypothesis, but it does not establish a single cause: three replacement cells cite different or no visible sources, and the provider may use undisclosed inputs. After a factual correction to an eligible public listing and a clarification on one owned page, a matched re-test records the rival in 5 of 18 cells (27.8%) and your brand in 4 of 18 (22.2%). Report those movements, then replicate; do not attribute all four changed cells to either edit.

Which false positives and alternative explanations matter?

Automated matching can count a common word, a former product name, a similarly named company, quoted prompt text, navigation, or a citation title as a brand appearance. An answer can list ten vendors without recommending any, recommend different products for different audiences, or cite a page that mentions both brands neutrally.

Even a correctly labelled change may reflect a provider experiment, index refresh, updated third-party page, current availability, account state, geography, personalisation, or stochastic output. Record those alternatives instead of turning temporal order into a causal claim.

• Verify brand aliases and competitor-domain ownership.
• Read the sentence around the name; do not score by string match alone.
• Distinguish an explicit recommendation from an example or exclusion.
• Compare exact conditions and retain an inconclusive label when they differ.
• Check whether the source or competitor page changed during the test window.

How do you choose the smallest supported fix?

Fix the layer for which evidence survived the disconfirming checks. If a verified access control blocks the exact relevant path, repair that control. If a public product fact is missing or outdated, correct it with attributable evidence. If an observed comparison source contains a factual omission, request a factual update; do not buy or fabricate endorsements.

Create or revise comparison content only when it serves documented buyer intent and can state verifiable criteria, trade-offs, and limitations. Keep structured data accurate and consistent with visible content, but do not present FAQ or other schema as an AI-citation lever: Google documents no special AI schema and does not guarantee appearance.

• Name the target URL or external record, owner, acceptance check, and ship date.
• Change one bounded layer where practical so the re-test remains interpretable.
• Preserve accurate counterevidence and do not copy a competitor's wording or claims.
• Pre-register the re-test window and exact cells before shipping.

How do you Prove whether the observed answer moved?

Repeat the same prompt wording, named engine and surface, repeat schedule, eligibility rule, locale, and account state as closely as possible. Compare the raw answers, visible URLs, your-brand state, competitor state, and unavailable counts before calculating a rate.

Describe a result as observed after the change, replicated, contradictory, or inconclusive. A stable panel and one bounded fix reduce ambiguity; they cannot freeze the provider, index, model, web, or third-party sources, so they do not prove that your edit caused the answer.

• Publish baseline and re-test numerators and denominators.
• Use percentage-point changes for rates.
• Keep unchanged and adverse outcomes in the proof record.
• Run additional scheduled repeats before promoting a one-window result.

What do official provider sources say?

Google says normal Search eligibility and snippet controls govern supporting-link eligibility in AI Overviews and AI Mode, with no additional technical requirements or special AI schema; eligibility does not guarantee appearance. Bing's February 2026 AI Performance announcement says citation count is not placement, ranking, authority, or page importance and that grounding-query data is sampled.

OpenAI documents OAI-SearchBot for search, GPTBot for potential training, and ChatGPT-User for user-triggered actions as separate controls. An access finding must identify which documented system and evidence it concerns; allowing one bot does not prove index inclusion, an AI impression, or a recommendation. The source records below were checked on 2026-09-01.

How does RankEcho hand the diagnosis from Find to Fix to Prove?

Use RankEcho's audit evidence to locate a repeated competitor outcome and inspect the exact answers and visible sources. Keep the proposed explanation labelled as a hypothesis until its checks are complete. Then send one supported, bounded change into Fix and retain the same cells for Prove.

The value of the workflow is the evidence trail: what was observed, what alternatives were considered, what changed, and what the same panel returned later. It does not guarantee that your brand will replace a competitor or that an observed movement will persist.

Sources reviewed

Provider eligibility and measurement claims below were checked against primary documentation. These records do not establish a universal selection formula, causation, or a guaranteed ranking, impression, recommendation, or citation.

Frequently asked questions