When is a brand the same brand? Entity resolution in AI answers

Entity resolution is the step that decides whether two names in AI answers refer to the same brand. It matters because every share, rate and ranking is built on counting brands, and counting goes wrong when one brand is counted as two or two brands as one.

What problem is it solving?

AI assistants write in natural language. A brand may appear in many forms:

• its full name, a short name, an abbreviation or a common misspelling;
• a name with or without a legal suffix;
• a product name, with no mention of the company;
• the company name, when the product was meant;
• a former name, after a rebrand;
• a name shared with something unrelated.

A human reader sorts these out without thinking. A measurement system has to do it explicitly, and consistently, thousands of times.

What goes wrong without it?

There are two opposite errors.

1. Splitting. The same brand is counted under several names. Each count is smaller than the true one, so the brand looks weaker than it is.
2. Merging. Different brands are counted as one. The merged figure belongs to neither, and a brand takes credit for another's recommendations.

Both are quiet. A report can look tidy and be wrong in a way no reader would spot. That is why entity resolution needs rules, not improvisation.

How does Citewake's published method do it?

In steps.

Find every brand in each answer. For each AI's answer, the reader returns every brand named, in order, with the exact words used, its role (recommended, a neutral mention, or negative), whether the order expresses preference, and any of its domains cited.

Check that the words exist. A brand is kept only if its name, alias or domain appears in the answer text and the quoted words are found verbatim. This proves the words are there. It doesn't prove the role or the match is right, which is why a separate accuracy check exists.

Resolve the entity. Brands are matched by domain first, then by alias. A brand in the method is a canonical entity with a name, aliases, products, a parent company and domains. Parent and product relationships are recorded: a product named in an answer counts for that product, and rolls up to its parent only where the parent is shown.

Hold ambiguous names. Ambiguous names are held, never guessed. A held name doesn't count for any brand until it is settled.

Why match by domain first?

A domain is a firmer identifier than a name. Many unrelated businesses share a name, but only one owns a given domain. When an answer cites a brand's domain, that points clearly to the brand. Names and aliases fill in where no domain appears.

What about products and parent companies?

A company may sell many products, and an answer may name only one. The method counts a named product for that product. The product is counted as itself, and the figure rolls up to the parent only where the parent is shown.

This keeps the figures honest. A product that gets recommended gets the recommendation, and the parent shows it where the parent is shown. It also prevents a large company from absorbing credit that belongs to one product line.

What about a name that means two things?

Some brand names are also ordinary words, places or other companies' names. When an answer says only that name, with no domain and no context that settles it, a careful system does not pick the likelier meaning and move on. It sets the case aside as unresolved. That costs a few data points and protects every figure from the quiet error of crediting the wrong company. Over time, the held cases also show where the alias list needs work: a new alias, a product to add, or a domain to record.

How do you know the matching is right?

By measuring it. The method uses a hand-labelled test set that deliberately includes ambiguous names, parent and product cases, brands that are easy to miss and incorrect additions. It reports entity accuracy per category, with confidence intervals, and sets bars a category must meet before anything is published. Why one reader for four AIs matters covers why one standard is applied to every assistant.

How can you do this on a small scale?

If you track a handful of brands yourself:

1. Make a list of each brand with its names, products, former names and domain.
2. Note any names shared with something else.
3. Decide in advance how you will count product names and parent names.
4. Review each answer against the list, and mark anything you can't settle as unresolved.
5. Keep the list, with a version and a date.

A simple spreadsheet is enough. The point is to decide the rules before looking at the results.

What should you ask of an AI visibility report?

• How does it decide that two names are one brand?
• What does it do with ambiguous names?
• Does it say how accurate the matching is?
• Can you see the answer behind a count?

Why every AI answer should be kept explains why the last question is possible only if the answers are stored. For the full description, see the methodology, published as a method ahead of the first edition in early 2027. To see how assistants describe your own brand, you can try the Citewake checker.