Why AI recommends your competitor instead of you

The question everyone asks first

When AI recommends a competitor, inspect the reasons it gives before deciding what to change. The answer may contain a wrong product fact. It may also describe a real advantage the competitor has for that buyer’s requirements.

A benchmark helps separate those cases. Its source links and stated rationale give your team evidence to review. They do not, by themselves, show why the model selected one source or prove that a content change will reverse the recommendation.

The 33,000 evaluations cover head-to-head comparisons and brand perception queries for companies in B2B software, scored with structured signal extraction across 39,000+ feature comparisons and 513 distinct features. Here’s what we actually found.

Models disagree with each other 48.6% of the time

This was the finding that surprised us most.

We looked at every prompt that was evaluated by at least two engines with a decisive outcome (win or loss, excluding ties). 2,358 prompts met that criteria. Of those, 1,145 produced different outcomes depending on which model answered.

48.6%. Nearly half.

Same company. Same competitor. Same prompt. Different engine, different winner. A buyer who asks ChatGPT gets told to go with Vendor A. A buyer who asks Gemini gets told to go with Vendor B. Both responses are confident. Both cite sources. Both sound authoritative. They just disagree.

This means a company’s competitive position in AI isn’t a single number. It’s four numbers, one per model, and they may tell opposite stories. The buyer’s impression depends on which app they happened to open.

Not all engines are equal

The disagreement isn’t random. Each engine has systematic biases that show up consistently across companies.

The spread is 8.3 percentage points on win rate between the most favorable engine (Gemini, 74.4%) and the toughest (Perplexity, 66.1%). That’s not noise. For a company running hundreds of buyer evaluations a month across models, 8 points is the difference between winning two-thirds of comparisons and winning three-quarters of them.

The absence numbers are even more striking. On Perplexity, companies are completely absent from 38.8% of brand perception evaluations. On Gemini, it’s 17.7%. A 21-point gap in whether the model even knows you exist.

What explains differences between engines?

The measurements above describe the historical sample. They do not isolate the effects of search providers, model versions, source freshness, or prompt handling. A difference in win rate should not become a claim that one engine always prefers a particular publisher or community.

Start with the captured answers. Compare the stated requirements and inspect the source passages attached to disputed claims. Record the interface, model when available, date, and search conditions so later reviews can distinguish a repeated pattern from a changed collection method.

Our earlier explanation assigned broad preferences to individual engines without evidence that separated those causes. We have removed those explanations. Use current source records for the question you are investigating.

The absence problem

In 26.6% of brand perception evaluations across all engines, the benchmarked company was completely absent from the AI’s response. Not misrepresented. Not mentioned with wrong facts. Just not there. The AI answered a question about the company’s category and didn’t include them. Same company, same category, same question: one engine includes you, another doesn’t.

Absence and inaccuracy call for different reviews. A wrong fact gives you a statement to verify. An omitted company requires checking the buyer’s requirements, product fit, and available coverage. Neither observation guarantees that a particular content edit will change a later answer.

28.6% of evaluations score below 50

This is the number that reframes the overall win rate.

The aggregate decisive win rate across all engines is 70.7%. That sounds fine. Most companies are winning most of their head-to-head comparisons most of the time. But the distribution has a long tail.

28.6% of all evaluations, nearly 1 in 3, produce an AI Presence Score below 50. 15.5% score below 30. In those evaluations, the AI is actively working against the company: wrong facts, competitor-biased framing, missing coverage, negative sentiment.

The 70.7% average hides these. A company can win 75% of its ChatGPT evaluations and lose 60% of its Perplexity evaluations. The blended number looks healthy. The per-engine number reveals a channel where buyers are being systematically steered away.

What should you inspect behind a recommendation?

Begin with the buying requirement. Read the answer’s stated reason for preferring a product, then check the factual statements that support that reason. A relevant difference in deployment or functionality may justify the recommendation. An outdated price may not.

Next inspect the cited passages. A competitor-owned comparison deserves review, but its ownership alone does not make it wrong or prove that it caused the loss. Preserve reliable attribution and leave the source unknown when the answer does not establish the connection.

Use the correction guide for a contradicted factual claim. If the comparison is accurate, decide whether your positioning needs more detail or whether the competing product is better suited to that particular requirement.

What this changes about monitoring

Monitoring one model tells you about the answers collected from that model. It does not establish what another service says about your company. Compare the systems relevant to your buyers and report each sample separately.

Investigate a weak result through the captured answer and its available evidence. Avoid attributing it to a search provider or community thread unless the record supports that connection.

There’s an irony in how we ended up exposing this. KnitKnot runs an MCP server, so you can connect Claude or ChatGPT directly to your workspace and ask the assistant itself how your AI presence changed this week: score trends per engine, mention rollups, competitor deep dives, the full picture. The same models that disagree about you become the interface for monitoring the disagreement.

Four models, four source ecosystems, four sets of buyer impressions. The recommendation your buyer gets depends on which one they ask. The question is whether you know what each of them says.

To investigate a repeated statement about your product, use the competitor claim monitoring guide. It covers the published claim, approved evidence, and its appearance in captured AI answers.