Why I Check Whether an Engine Can Name the Brand

I check whether an answer engine can resolve the company name, the product, the press, and the market claims into one entity before I spend to win a citation. If it cannot, the category can be discussed all day and the company never enters the answer. Entity resolution rate is the share of prompts on which the model identifies the brand as a single entity and attaches the right claims, products, and sources to it. Citations are downstream of that identification. A citation chart, read before the identification check, is a chart of other companies.

The definition and the formula live on Entity Resolution Rate: Definition and Formula | MR Research. That page is the deeper resource. This essay is how I would use it as a founder: when I would stop a citation program and clean up identity, when I would do the opposite, and which figures I would allow into that decision.

What the rate is measuring

In enterprise data work, entity resolution links records to the same real-world thing. Answer engines face the same problem when a prompt about a category should or should not resolve to a company. The research literature still treats broken identity linkage as a way to corrupt every downstream decision (MERAI, arXiv). I would use that as the reason a naming mess is not a style problem. It is an input problem.

The disagreement across engines is now measured. A 2026 brand tracker running 8,400 prompts across four engines found the engines agreed on the top-cited brand only 34% of the time, and only 21.4% of the time on comparison queries. For nearly one in five queries, no two engines cited the same brand at the top (Visionary Marketing, AI search visibility statistics, 2026). I would not call a win in one engine a category position while the next engine resolves the same prompt to a competitor. The 34% and the 21.4% are why.

Metric	What I would let it say	What it does not say	When I would look at it
Share of voice	How often the name appears in a market conversation	Whether the mention is the right entity	After the name is stable
Share of citation	How often the brand is cited inside answers	Whether the engine could resolve us before it chose	After resolution is strong
Entity resolution rate	How often engines identify the brand as one entity and attach the right claims	A revenue outcome	Before either of the other two is allowed to move spend

If I only measure share of citation, I am reading the output of a system and skipping the question of whether the system can see the input.

What a weak rate does to eligibility

A weak rate means the brand can lose before the citation contest starts. If the engine cannot resolve the company, it cannot confidently attach a third-party validation, a product description, or a category fit to the right entity. The company is ineligible in the moments that matter, not invisible on every keyword.

Concentration makes the eligibility expensive. One analysis found the top-cited brand in a sector taking an average of 31.4% of that sector's brand citations, and the top three combined taking 64.7% (Visionary Marketing). The exact share moves by sector. The shape is why I care: if we are not resolving into the set the engine can name with confidence, the citation share accrues to brands it can name. More pages do not fix that. More pages with inconsistent names, an unclear product architecture, or shifting category language give the model more tokens and less certainty.

A 2026 entity-matching benchmark of 755,540 labeled pairs across 293 sources and 31 countries found that matching quality still depends on noise, ambiguity, and representation, even with current models (OpenSanctions Pairs, via arXiv). A company can appear as a parent brand, a short product name, a nickname, an old category, or a press shorthand. When those strings do not converge, the system does what weak matchers do. It prefers the clearer alternative.

Why I would treat this as a commercial system, not a data-cleanup ticket

Once discovery runs through answers, identity quality decides who is legible enough to recommend. The infrastructure papers are the scale argument. MERAI, a 2025 pipeline, reported that common tooling failed beyond 2 million records while its pipeline handled sets up to 15.7 million records (arXiv). A separate study found up to 150% higher accuracy and a 10% increase in F-measure from stronger in-context clustering, while cutting API calls by up to 5 times (arXiv). I would not copy those percentages onto a marketing dashboard. I would use them for one claim: serious systems spend real effort on identity because everything downstream depends on it. An answer engine has to decide whether the pricing page, a quoted executive, a magazine mention, the product page, and the category term are one thing.

The commercial measurements point the same direction. Entity-optimized brands see up to 70% more accurate AI-generated descriptions than competitors that were not optimized, in the account I am using. Sites with complete Organization, Brand, and AboutPage schema were cited 3 times as often in AI shopping results across 150 ecommerce audits (Onely). I would brief the 3x as a schema result from those 150 audits, not as a promise that adding three schema types triples our citations. The direction is what I would fund: complete, consistent identity markup after the words already agree.

Where this sits relative to the work founders already buy

GEO and AEO format and distribute pages so a machine can lift them. They do not, by themselves, repair an identity. If the entity layer is broken, distribution amplifies a fragmented record. I would track whether systems resolve us, and only then watch which citations follow.

Discipline	The optimization	The close I would accept	Scope
SEO	Ranking systems	A top position on a results page	Technical work and content
GEO	Generative engines	Cited in an answer	Format and distribution
AEO	Answer boxes	Selected as the direct answer	Structured pages
Digital PR	Human editors	A placement	Outreach and story
Machine Relations	Answer engines	Resolved, and then cited	Authority, entity, citation, distribution, measurement

The operational cleanup is dull, which is why it loses budget fights to a campaign. I would still do it in this order. Keep the brand name, the product name, and the category label stable. Describe the company in the words the market uses. Require third-party articles to use the same descriptors. Point executive profiles back at the company. Put the strongest proof in more than one corroborating source.

Third-party description is the load-bearing piece. One analysis found that 70% to 80% of the public mention footprint for brands that stay visible in AI answers came from third-party sources, most of them on domains the brand did not own (BrandMentions). I would rather have those outside sentences agree than have another homepage variant.

The trap during growth is ambiguity we create. Two repositionings in a year, a renamed product, contradictory homepage copy, an old review-site description, a database summary, a launch article, and an interview that describes a different company. The answer then feels random. It is unresolved. I would freeze the strings before I scaled the content.

There is also a clock on the citation once it lands. The same 8,400-prompt tracker found a citation persisting at the same brand for an average of 41 days before it drifted. On Claude the median was 67 days. On Perplexity it was 18 (Visionary Marketing). I would not treat 41 days as a guarantee. I would treat it as a reason to resolve the entity before the window, because a citation that attaches to a muddy name is a citation we may spend the next month correcting. The brands that resolve first are the brands the persistence can compound. The brands that skip the check keep asking why a competitor is the name in the answer.

I coined Machine Relations in 2024 to have a name for that order: identity, then authority, then recommendation. The stack I would hand a team that wants to start at the citation is Machine Relations. The metric I would make them define before they quote a percentage is entity resolution rate.

How I would run the check before the next budget meeting

I would take the buyer prompts we already use for citations and add a second score on the same runs, so the identity check does not become a separate research project. For each prompt I would mark one of four outcomes. The engine names us and attaches the right product and the right claim. The engine names us and attaches the wrong product, the old category, or a fact we do not make. The engine describes the category and names only competitors. The engines disagree on the top name. The 8,400-prompt tracker is the outside version of that fourth outcome: agreement on the top-cited brand 34% of the time, 21.4% on comparison queries, and nearly one query in five where no two engines share a top brand. Our sample will be smaller. I would still publish the four counts with the prompt list, and I would not publish a citation share next to them until the first outcome is the common one.

The concentration figures tell me why the order matters. A top brand taking 31.4% of a sector's brand citations, and the top three taking 64.7%, means the engine's confident names collect most of the mentions. Cleaning our schema while the name still splits across a parent, a product, a nickname, and a stale category is decorating a record the matcher will not assemble. The OpenSanctions benchmark, 755,540 pairs across 293 sources and 31 countries, is the research statement of that limit: better models do not erase messy attributes. MERAI's report that ordinary tooling failed past 2 million records, while a purpose-built pipeline reached 15.7 million, and the separate finding of up to 150% higher accuracy, a 10% F-measure gain, and up to 5 times fewer API calls, are not numbers I would paste into a board slide as marketing results. They are the reason I would stop treating identity as a ticket for someone to clean up after the campaign.

Onely's 150 ecommerce audits, a 3x citation gap for sites with complete Organization, Brand, and AboutPage schema, and the "up to 70%" more accurate descriptions for entity-optimized brands, come after the words agree. I would complete the schema when the homepage, the bios, and the third-party sentences use one name and one category. I would not add schema that points at a profile or a description we have since abandoned. BrandMentions' 70% to 80% third-party share of the mention footprint, mostly on domains the brand does not own, is why the outside sentences are part of the check and not a nice-to-have. If those sentences describe a company we used to be, the rate will stay low no matter how clean the homepage is.

The persistence numbers change what I do after a good week. An average of 41 days on the same brand, a Claude median of 67, a Perplexity median of 18, means a correct resolution is worth reinforcing inside the month, and a Perplexity hit is a poor thing to celebrate alone. I would schedule the corroborating edit or the placement while the citation is still attached, and I would re-score the four outcomes before I called the spike a trend. A citation stuck to the wrong product for 41 days is a month of teaching the engine the wrong assembly. That is the failure I would rather spend to prevent.

The decision

I would not scale citation spend while a sample of buyer prompts shows the engines attaching our claims to the wrong entity, splitting us across names, or describing the category without naming us. I would spend that cycle on one canonical name, one category label, aligned bios, and third-party sources that repeat the same description. If the resolution rate is high and the citations are still weak, I would switch budgets. The machine knows who we are and does not yet have enough authority to recommend us. That is a distribution problem, and it deserves the spend the identity problem no longer needs.

Scope: companies tracking AI visibility that have been reporting mentions or citations without an identity check. The agreement figures, the 31.4% and 64.7% concentration, and the 41, 67, and 18 day persistence numbers are from the 8,400-prompt tracker. The schema 3x is from 150 ecommerce audits. The 70% to 80% third-party footprint is from the BrandMentions analysis. None of these is a formula that predicts our revenue. The research page owns the formula for the rate. I am deciding when to use it.

What would change my mind: a disclosed multi-engine panel, on the order of those 8,400 prompts, in which top-brand agreement rose far above 34% and comparison-query agreement far above 21.4%, and in which messy public naming stopped changing who was cited. I would then treat identity as maintenance and buy citations first. I would also skip ahead if our own prompt sample already resolved us cleanly and the only gap was authority. I would not skip ahead because a dashboard showed a citation spike on a single day. The persistence numbers say a spike can be a 18-day Perplexity event or a 67-day Claude event. I would want to know which, and I would want to know the name it attached to.

FAQ

What am I checking, in one sentence?

Whether, on the prompts that match our buyers, the engine identifies the company as one entity and hangs the right claims, products, and sources on it. The share of those prompts is the entity resolution rate. The formula and the boundary conditions belong to the research page this essay points at. I use the rate as the gate in front of citation spend.

How is that different from share of citation?

Share of citation tells me how often we are cited after identification has succeeded. Entity resolution rate tells me whether identification succeeded. A weak identity makes the citation comparison a false read, because the engine is ranking brands that fully entered the set. I would not reallocate budget on the citation number until the identity number is strong enough to trust.

What produces a low rate?

Inconsistent names, fragmented product descriptions, thin third-party corroboration, and mixed category language. The OpenSanctions benchmark, 755,540 pairs, 293 sources, 31 countries, is the research version of the same fact: noise and ambiguity still degrade matching. A startup that repositioned twice, renamed a product, and left an old database blurb in place has built that noise on purpose.

What do I do when the engines cannot name us, and how long does a citation last if they can?

I fix the strings and the corroboration before I buy more citations: one name, one category, bios that point at the company, and outside sources that describe us the same way. Onely's 150 ecommerce audits are why I would also complete Organization, Brand, and AboutPage schema once the words agree: those sites were cited 3 times as often. I would not expect the schema to repair contradictory prose. When a citation does lock on, the 8,400-prompt tracker saw it persist an average of 41 days, a median of 67 on Claude and 18 on Perplexity. I would use the longer window to reinforce a correct description and the shorter one as a reason not to celebrate a single Perplexity hit.