Why Naming Conventions May Fix AI's Biggest Problem

Ask an AI about a business, a person, or a product, and you will get a confident answer. The trouble is that confidence is not the same as knowledge. A language model does not look up names in a registry. It predicts them, the same way it predicts every other word.

Researchers have measured the damage. In a 2025 Northwestern University study, three widely used AI tools were tested on a reporting task grounded in 300 real documents. Thirty percent of the model outputs contained at least one hallucination, and most errors were not invented facts but attribution failures: models added unsupported characterizations of sources and turned attributed opinions into general statements. The authors concluded that newsroom AI needs architectures that enforce accurate attribution rather than optimize for fluency.

A separate line of research tested whether models can attribute short passages of famous books to their authors. Even the best model hallucinated badly on some books, and the researchers found that prediction accuracy tracked how often the title appeared on Wikipedia. In other words, the famous get recognized and everyone else gets guessed at.

A 2026 analysis of brand confusion in AI models lays out the mechanism. Large language models maintain no canonical registry of companies, products, or organizations. When a name appears, the model does not verify its real-world identity. It selects the statistically most plausible interpretation in context.

Entity resolution inside the model runs as a sequence: detect a possible named entity, parse the surrounding context, pick the most likely candidate, then generate the answer as if that pick were correct. There is no validation step. A low-confidence guess still produces a fluent, authoritative response.

Ambiguity is everywhere: brand names collide with common nouns, technical terms, places, and other brands. Research summarized in the Stanford AI Index 2024 indicates that a significant share of brand-related model outputs contain entity misidentification or conflation.

This is the huge problem: as AI becomes the default interface for discovery, the system deciding what things are called is a probability engine with no ground truth.

Here is the thesis, stated as opinion: the fix is not a smarter model. It is more disciplined naming by the rest of us.

Models resolve entities by statistical weight. A name that appears in one canonical form, consistently, across your website, your listings, your structured data, your press mentions, and your transcripts gives the model a mountain of matching text to draw from for every query. A name that appears in five variations gives it five weak signals competing with each other.

We have seen this mechanism work. A 5WPR study of 95 podcast hosts across five AI answer engines found that citation share does not track audience size. Hosts who publish consistent, long-form, single-topic content get cited more, because the model has more matching text per query. Names work the same way: consistency is the signal.

A naming convention is a promise you make to every future model: this thing is always called exactly this.

First, your public identity. Pick one canonical name for your business, your products, and your key people. Use it identically everywhere: website, Google Business Profile, LinkedIn, directories, and structured data. Every variation is a vote for a different entity.

Second, your data and schemas. AI agents act on field names, file names, and labels. Ambiguous names like final_v2 or data_new are how AI agents confidently operate on the wrong dataset. In our own work, client material is pseudonymized with consistent labels before any AI processing, because consistent naming is what keeps identities from leaking across documents.

Third, your content library. One topic, one canonical title pattern, one naming scheme for files and episodes. The models are counting. Give them something clean to count.

Honesty requires the limit. Consistent naming raises the odds of correct resolution. It does not give the model a ground truth it lacks. A disciplined name cannot fix a model that never saw your business, and it cannot stop confident fabrication. Verification still belongs to a human. Naming conventions are the cheapest reliability upgrade available, not a cure.

Five steps, no new software required:

• List every public name for your business, products, and people. Count the variations.
• Choose one canonical form for each. Write it down where your whole team can see it.
• Put the canonical names in your website's structured data.
• Apply the same discipline to the files, fields, and datasets your AI tools touch.
• Then watch: the next time an AI describes you, check whether it used your name or its guess.