Common-Word Brand Names and AI Mention False Positives

If your company is called Pulse, Scout, Beacon or Craft, your AI visibility number is probably too high — and the fix is entity disambiguation, not more prompts. An independent test of 12 AI visibility platforms found that 9% of reported mentions were false positives, measured against 600 manually verified prompt-answer pairs across five surfaces in May and June 2026. That 9% is the average across all brands tested. For a brand whose name is an ordinary English word, published audits put the error rate far higher: in one case, a startup's 30 reported mentions turned out to be 28 references to a fictional character who happened to share the name.

Two mentions were real. The dashboard also credited the company with 600K+ estimated audience reach, and nothing on the screen looked broken — because nothing was broken. The tool found the string it was told to find. That is the whole problem in one sentence.

Why "Pulse" is a measurement problem, not a branding problem

Start with how crowded the name space actually is. Wikipedia's disambiguation page for Pulse lists dozens of distinct entities — roughly 20 in film, television and games, about 15 in music, eight in technology and seven in transportation — including the Fiat Pulse, a 2025 Netflix medical drama, the LinkedIn-acquired Pulse news app and the Livescribe Pulse Smartpen. Scout is comparably crowded: a global youth movement, a long list of military units and aircraft, the International Scout off-roader, Volkswagen's Scout Motors.

Every one of those is a thing an AI engine might mention in an answer. A tracker doing plain string matching counts all of them as you.

And the name space keeps growing without your permission. OpenAI launched ChatGPT Pulse on 26 September 2025, shipped ChatGPT Atlas on 21 October 2025, and retired Atlas on 9 August 2026 — two days ago — folding it into ChatGPT and Codex. A substring-matching tracker gave a company called Pulse or Atlas an unearned spike after each launch and will hand it an unearned decline after the shutdown. You spend the quarter explaining a trend line that describes someone else's product roadmap.

Which kills the most common defence of a noisy metric: constant bias is survivable, but this bias isn't constant.

The three ways a false positive gets into your dashboard

Practitioner write-ups on brand mention tracking name three distinct failure classes, and each needs a different fix.

Name collision is the obvious one: the engine is talking about a person, a place, a feature, a car or another company that shares your name. Category-term collision is sneakier — your brand name is the category word, so the engine's reference to the concept registers as branded. Phantom inclusion is self-inflicted: your monitoring prompt contains your own brand name, so the answer mentions you because you asked it to. That measures prompted recall, not organic visibility — and it's the easiest of the three to fix.

Underneath all three sits one mechanical cause. Substring matching without word-boundary enforcement is the largest single source of false positives in mention detection: a brand called Craft matches inside "aircraft," "handcrafted" and "craftsmanship." The independent test above traces its false positives to the same root — "when an AI answer references a company name that appears in multiple contexts, weaker detection layers guess rather than verify."

The reason this survives quarter after quarter is that it never announces itself. As that audit puts it: "A tool tracking the wrong entity produces plausible, populated, professional-looking reports. There's no error state." A crawler misconfiguration throws an error. A false positive throws a chart.

It is also the worse of the two possible errors. A false negative undercounts you, which is annoying and correctable. A false positive "tells you a problem is solved when it isn't and pollutes every downstream metric" — sentiment, prominence, share of answers. If 28 of your 30 mentions are fan posts about a fictional character, your sentiment score is measuring how much people liked that character.

The same name breaks your visibility in the other direction

Here's the part that catches teams off guard: a common-word brand has two problems running at once, and they point opposite ways.

The first is the measurement bug above: your tracker counts other entities as you, so your reported number is too high. The second is a visibility bug: when a buyer asks about you, the engine may answer about someone else. Language models resolve names to entities, not strings, binding an ambiguous name to whichever node carries the strongest training and knowledge-graph signal — which, for a small brand sharing a name with a Netflix series, is not you.

That's worth distinguishing from hallucination, because the fixes differ. Hallucination is the right entity with invented facts, fixed with authoritative content; disambiguation failure is the wrong entity with accurate facts about them, fixed with entity signals. One agency auditing its own name across four engines found 11 factual errors — ChatGPT placed it in San Diego, founded 2012; Gemini called it an AI image generator — when it is an AI product studio in Makati, founded in 2024.

Engines also disagree about which signals to trust — ChatGPT leans on training data, Perplexity on live retrieval plus schema, Gemini on the Google Knowledge Graph, Claude on long-form documentation. So a common-word brand is routinely resolved correctly on one engine and wrongly on another, which is why a per-engine breakdown beats a blended number. A free AI Visibility Check is the five-minute test: run one query across all four engines and read the answers rather than the score. Are they describing your company, or your namesake?

What entity disambiguation actually looks like inside a tracker

A serious detection stack runs four layers, and the order is the point: an LLM judge bolted onto a bad string matcher inherits the bad matches.

1. Normalise — lowercase, Unicode NFKC, strip diacritics, tokenise on word boundaries. This also catches the false-negative twin of the problem: the same brand written Loo.koo.mas and Lookoomas is missed entirely by exact matching.
2. Exact alias gazetteer — canonical name, legal name, product sub-brands, known misspellings.
3. Conservative fuzzy match — token-set ratio around 80–85, tightened to 90+ for short names, because short names are where fuzzy matching does the most damage.
4. A judge with an ambiguity hint — the disambiguation model gets your canonical name, aliases, competitor list and an explicit warning ("Pulse is also a Fiat model, a Netflix series and a LinkedIn product"), so it can catch paraphrase and reject same-name confusion.

Then comes the layer that actually changes your number: a corroboration gate. A mention counts only when an anchor is present — your domain is cited, a product term appears in the same sentence, or the answer clearly describes your positioning. One published pipeline scores own-domain-cited at 1.00 and competitor-co-occurrence at 0.85 (both counted), against name-with-no-anchor at 0.50 and judge-disagrees at 0.40 (both excluded).

What does gating cost? On a deliberately ambiguous name, 15 grounded answers were analysed, 10 contained a brand signal, and 4 survived confirmation. Presence with the gate: 26.7%. Presence if every named answer counted: 66.7%. The gate cut the headline number by two-thirds — and the smaller number was the true one.

Vendors know this is the hard part. Semrush documents contextual brand extraction rather than simple text matching, using "Tesla (the EV company), Nikola Tesla (the scientist), and Nikola Tesla Airport in Belgrade" as its worked example, while conceding that "no platform can provide exact numbers on visibility." Otterly is credited with entity definition controls. The capability exists; it just isn't standard, which is why you have to ask.

The precision audit: one afternoon, one honest number

You don't need a vendor's permission to find out how bad your own case is. Precision — "the proportion of all the model's positive classifications that are actually positive," or TP / (TP + FP) — is the metric to audit here, not accuracy.

Freeze the prompt set and strip your brand name out of every prompt, so you stop manufacturing phantom inclusions. Export the raw answer text, not just the flags; if a tool won't show the text behind a flag, you cannot audit it. Sample 50–100 flagged answers, have two people label each one genuine or not, and compute precision as genuine ÷ flagged — checking the labelling itself with Cohen's κ between the labellers (above 0.6 is workable, above 0.8 is strong).

Now do the arithmetic that changes the meeting. Adjusted mention rate = reported mention rate × precision. Say your dashboard shows 48%. At the cross-tool average false-positive rate of 9%, your real number is about 43.7% — irritating, not fatal. At the precision from that ambiguous-name demo (4 genuine of 10 flagged, so 0.40), the same 48% is really about 19%. That's the gap between "we lead the category" and "we're a niche brand performing normally."

Then sanity-check the result against category norms. A study of 100K+ prompt responses across 100+ brands found global household names appear in 73% of relevant answers, mid-market brands in 44%, and niche brands in just 11%. If you are a Series A company reporting 48%, the likeliest explanation is a measurement problem, not a marketing win.

Two disciplines keep that adjusted number defensible. Size the sample: roughly 228 prompts buys a ±5pp margin of error, under 50 prompts the interval is too wide to act on, and three runs per prompt is the floor because most variance lives at the run level. And treat the result as an estimate rather than a value — visibility metrics behave as sample estimators of a response distribution, which is why a headline score belongs inside a rolling window with a confidence band. Our guide to choosing the 25 prompts you track covers the composition side of the same instrument.

One thing most teams skip: run the same precision check on your competitor set. If a rival has the ambiguous name, their inflated count quietly deflates your share of answers, which distorts every competitive share-of-answers analysis you build on top of it.

Fixing the entity so the number stops lying both ways

Auditing precision fixes your reporting. Fixing the entity fixes the underlying visibility — and it's the same work either way.

Pick one canonical name and use it identically everywhere. Ship sitewide Organization schema with name, legalName, alternateName and a sameAs array pointing at LinkedIn, Crunchbase, GitHub and Wikidata, and include disambiguatingDescription — a standard schema.org property for "a short description of the item used to disambiguate from other, similar items." Create the Wikidata item too, with instance-of and industry set correctly; it's the strongest sameAs target, because it feeds the Google Knowledge Graph, Wikipedia and downstream training pipelines. Then open your entity home page with a flat declarative line — "Pulse is a [category] for [audience], based in [place], founded [year]" — and where a namesake dominates, say so outright: not the Fiat model or the Netflix series.

Set expectations on timing. Schema, Wikidata and third-party profile changes take weeks to propagate as engines re-crawl their sources, so re-measuring a fortnight later reads noise, not progress. Plan the recheck for the following quarter, and use the interval to rebuild your alias, anchor and negative-term config so the before-and-after comparison is at least measuring the same instrument.

What to ask a vendor before you trust the number

Five questions separate a tool that disambiguates from one that greps:

• Do you enforce word boundaries, and normalise case, punctuation and diacritics before matching?
• Can I supply aliases and negative terms — the industry vocabulary of my best-known namesake?
• Is there a disambiguation judge, and what is its measured agreement with human labels?
• Do you expose the raw answer text so I can verify a flagged mention myself?
• Is a mention counted once per answer or once per occurrence? (Semrush counts per unique prompt; per-occurrence counting inflates differently.)

There's a structural argument for a composite score hiding in that last question. Take a published formula like Geoptimizer's EngineScore = 100 × (0.35·MentionRate + 0.25·CitationRate + 0.20·Prominence + 0.20·Sentiment). Mention rate, prominence and sentiment all depend on having located the right entity in the text — 75% of the score resting on entity resolution — while citation rate is anchored to a domain match that cannot be ambiguous. If your name is a common word, that citation quarter is your most trustworthy signal. You can only reason this way when the weights are published, which is the case we make in the evidence behind our scoring formula.

And be straight about the headline figure: the 9% comes from one test by one agency, with the platforms identified only by tier, so nobody can replicate the per-tool results. It's an order-of-magnitude baseline. Your own precision number is the one that matters.

FAQ

How do I know if my AI visibility tool is counting the wrong brand? Export the raw answers behind 50 flagged mentions and read them. If the tool won't show you answer text, that's your answer. Precision = genuine ÷ flagged; anything under about 0.8 means the reported mention rate needs adjusting before it goes in a deck.

My brand name is a common word. Should I just track citations instead of mentions? Citations resist false positives — a domain match can't be ambiguous — but only 8 of the 12 platforms tested recorded which sources the AI cited, and citations appear in a minority of answers. A citation-only metric trades a precision problem for a recall problem. Track both, and weight citations more heavily when your name is ambiguous.

Doesn't a constant error rate still give me valid trends? It would, if the error rate were constant. For a common-word brand it moves with the news cycle and with unrelated product launches and shutdowns that share your name — ChatGPT Pulse's launch, Atlas's retirement. Non-stationary bias corrupts deltas, not just levels.

Won't an LLM judge solve this? Partly. Reported judge-to-human agreement in 2026 practice tops out around 92% with κ ≈ 0.85, judges drift over time, and they carry the same popularity prior that caused the problem — ask a judge "is this the company Pulse?" and it leans toward the famous Pulse. Calibrate against a human-labelled sample and re-sample monthly.

Are big brands immune to this? No, they fail differently. A June 2026 paper argues that aggregate mention rate is insufficient precisely because entities have systematically different error profiles, naming a "Brand Hallucination Paradox" in which model familiarity with large brands "creates stronger surfaces for plausible but incorrect completions". The study's sample is narrow — 100 Hungarian B2B entities — so read the direction, not the percentages.

The number you report should be the number you'd defend

Prompt count is the easiest thing for a tool to sell and the least interesting thing about a visibility score. If your brand name is an ordinary word, what decides whether your number means anything is entity disambiguation: word boundaries, aliases, negative terms, a judge that knows your namesake, and a gate that would rather undercount you than flatter you. Everything else is arithmetic on top of a guess.

So run the precision audit this quarter, report the adjusted number internally, and re-measure only after your entity fixes have had weeks to propagate. For a starting read, run a free AI Visibility Check across ChatGPT, Gemini, Claude and Grok and read the answers before the score — then compare the plans when you want to track it continuously with a formula you can audit line by line.