# What does an AI comparison win rate mean when answers tie?

> A win rate can hide two different problems: AI prefers a competitor, or it declines to choose. Keep wins, ties, and losses visible before changing your content.

- Author: KnitKnot
- Published: 2026-10-07
- Canonical: https://knitknot.ai/learn/what-does-ai-comparison-win-rate-mean/
- Publisher: KnitKnot, the AI Competitive Positioning platform (https://knitknot.ai)

---

## Why can a good comparison record have a low win rate?

Because a tie can count in the total. Suppose you capture eight answers to buyer questions comparing your product with a rival. Four favor you, three name no winner, and one favors the rival. These are example counts, not a KnitKnot customer result.

| Outcome | Answers | What it means in this example |
| --- | ---: | --- |
| Win | 4 | AI favors you |
| Tie | 3 | AI compares the vendors but does not choose one |
| Loss | 1 | AI favors the rival |

With ties included, your **win rate is 4 ÷ 8 = 50%**. Among answers that picked a winner, your **decided win rate is 4 ÷ 5 = 80%**. Neither number is wrong. They answer different questions.

The gap between them tells you to read the tied answers. If you only saw 50%, you might assume the rival won half the comparisons. It won one of eight. If you only saw 80%, you might miss that three answers gave the buyer no clear recommendation.

## What should count as a tie?

A tie is a comparison outcome where the answer does not favor either vendor overall. That can happen because the answer says the choice depends on a buyer's needs, because each vendor wins a different criterion, or because it lacks enough evidence to decide. Those are different stories. Read the captured answer before assigning a cause.

Keep **not compared** separate. A question that did not compare vendors, or an answer without a usable comparison outcome, should not be quietly treated as a tie. Ask your measurement provider how it handles those answers and whether the raw response is available.

## Which rate belongs in a report?

Show the **wins, ties, and losses together**, then label each percentage with its denominator:

- **Win rate:** wins ÷ (wins + ties + losses). This tells you how often you were the clear winner among measured head-to-head outcomes.
- **Decided win rate:** wins ÷ (wins + losses). This tells you how often you won when AI picked a winner. It is undefined when there are no wins or losses.
- **Tie share:** ties ÷ (wins + ties + losses). This shows how often the answer left the comparison unresolved.

In the example above, tie share is 3 ÷ 8 = 37.5%. Keep the count next to the percentage. A change from one win in two answers to two wins in three answers is too small a base to present as a settled shift in buyer perception.

Other products may define “win rate” differently. Ask for the formula rather than assuming their percentages are directly comparable. KnitKnot's [metrics reference](/docs/metrics-reference/) defines both rates used in its Head-to-head table.

## What do you investigate next?

Open a few ties and losses from the **same competitor, buyer question, and AI engine**. Look for the criterion that shaped each answer: price, feature scope, integration, security, implementation effort, or something else. Then check the statement against your approved product facts and the sources the answer actually exposes.

If the answer gets a material fact wrong, gather the exact quote and a current source before preparing a correction. If it accurately describes a limitation, the next move may be a clearer positioning page or a product decision. If it simply says “it depends,” see whether the comparison names the buyer conditions that would make your product the better fit. A blanket “we win” page would not answer that question.

After a change, rerun a comparable set of questions. Keep the prompts, engines, and outcome rules steady enough to see what changed, and read the answers again. A later improvement is an observation; it does not prove that one page caused it. For the product workflow, see [Read your report](/docs/read-your-report/) and [Investigate a claim](/docs/investigate-a-claim/).