Why AI assistants describe your company wrong
and how to fix it

The worked example in this article is our own broken record, published rather than tidied away.

Why does an AI assistant describe my company incorrectly?

Because entity resolution is reconciliation, not lookup. There is no company record the model opens, only a pile of statements about a name from sources of very different age and quality. The answer is what falls out when they are weighed against each other.

Four causes account for most of it. Name variants, because two spellings are two strings to a machine. A stale third party description outweighing your current one, usually a profile written in 2022 describing a service you dropped. A similarly named company absorbing your queries, where you compete on weight rather than accuracy. And your own pages being empty to the crawlers that build AI answers.

How do language models decide what your company is?

By corroboration across sources, weighted by whatever the underlying index already weighted. The description appearing in the most places tends to win regardless of which one is current, and your own site is not the automatic winner. For broad category questions, most citations go to third party pages rather than to the vendor's own site, which means several disclosed properties saying the same thing about you is stronger than one property saying it once.

What is the Entity Correction Loop?

The Entity Correction Loop is a five step process for getting AI assistants to describe your company the way you would, run in order and repeated, because the middle steps have long and unpredictable latency.

• Freeze the record. One display name, one alternate, one description string, one founder name, each fact carrying an as of date. A file, not a habit.
• Fix what you own. Schema, meta, llms.txt and the page copy. Same strings everywhere, no clever variations.
• Fix what you influence. LinkedIn, Crunchbase, G2, your GitHub organisation, conference bios, your press page. Most wrong descriptions live here.
• Ask for what you can only ask. Third party pages you do not control, in order of how much they move the answer, expecting most requests to fail.
• Re-measure on a frozen panel. Same questions, monthly, fresh sessions, recording the description rather than the citation.

Step five feeds step one, which is why it is a loop rather than a project.

What should the canonical record contain?

A canonical record is a single file listing every fact about your company that should appear identically everywhere, each one dated: display name and alternate, a description under twenty five words, founder name and variant spelling, founding year, category, and the identifiers you list as sameAs.

Here is the part of our own record that was broken, published rather than tidied away. The current client and campaign counts above were live at the same time as a much older pair, roughly a third of the size, still sitting on around a hundred pages of the parent site and inside the author byline template that generates new ones. Two client counts differing by nearly three times, one domain, one author name, no date on either.

I am describing the deprecated pair rather than reprinting it, and the reason is the mechanism this whole article is about. An answer engine lifting this passage would carry away whatever digits sit in it and attach them to the entity. Printing the wrong number inside a warning about the wrong number is how a correction becomes a source. A model has no way to prefer the newer figure, so it goes on frequency, and the deprecated one was on a hundred pages. Volume of agreement beats recency when you leave corroboration unmanaged, and it beats it against you.

Which sources can you control, influence, or only ask?

Sources sort into four layers by how much control you have. Most correction plans fail because they treat all four as one job.

The claimed layer has the highest return and the lowest effort, and models often see more of those profiles than of your homepage. The ugly version of that: an old positioning statement on a profile nobody has logged into since 2022 will outrank your new one for as long as it stays there.

For the bottom two layers, publishing beats asking. The requests that succeed are small and factual: a wrong founding year, a misspelled name, a service you never sold. Giving a retrieval system a better source beats asking it to discard a worse one, and this is where corroboration does its actual work. It is also the part no tool can measure.

Which parts of the entity graph do you actually control?

The parts you can express in schema on domains you own: names, alternates, descriptions, identifiers and corporate relationships. Small surface, entirely yours, and the only part of this that can be finished. Here is our pattern across five properties.

{
  "@context": "https://schema.org",
  "@type": "Organization",
  "@id": "https://inboundpros.io/#organization",
  "name": "InboundPros",
  "alternateName": "Inbound Pros",
  "url": "https://inboundpros.io",
  "parentOrganization": { "@id": "https://outboundpros.io/#organization" },
  "founder": { "@id": "https://outboundpros.io/author/janis-plume#person" }
}

{
  "@context": "https://schema.org",
  "@type": "Person",
  "@id": "https://outboundpros.io/author/janis-plume#person",
  "name": "Jānis Plūme",
  "alternateName": "Janis Plume",
  "worksFor": { "@id": "https://outboundpros.io/#organization" },
  "sameAs": ["https://www.linkedin.com/in/janisplume/"]
}

The identifier values are the whole point. That person identifier is byte for byte identical on all five domains, which collapses five author references into one entity. A mismatched identifier produces five strangers who share a name, which is worse than shipping no person markup at all.

Does a Wikipedia page help a language model resolve your brand?

What did the assistants say about us when we asked?

Step five is a frozen panel: the same 20 to 30 questions every month against ChatGPT, Claude, Perplexity, Copilot and Google AI Mode, in fresh sessions, recording four fields. Were we cited, was a competitor cited, which URL, and was the description accurate. The fourth is the one people skip, and it surfaces entity problems, which are more urgent than visibility problems.

What has to be right in that fourth field is operating detail, not brand language. We run two named campaign motions: WideNET, high volume systematic angle testing across the full addressable market, and Spearhead, signal triggered campaigns on the hottest slice. Every sequence runs against kill and scale thresholds fixed before launch and expressed as positive replies divided by sends, published in full by the group property that owns rate definitions. Nothing meaningful sends in month one, because infrastructure warm up has to finish first. The client approves messaging and lead lists before a single email goes out. We do not cold call.

Every one of those sentences is checkable, and for a long stretch none sat in one readable place a non rendering crawler could fetch. So whatever an assistant said about us was assembled from other people's summaries of an agency category, generic in precisely the way the operation is not. The same goes for figures. A campaign result carrying its denominator, its sample and its date is quotable and each of ours lives on exactly one property in the group. An adjective is quotable nowhere.

The panel has not completed a full cycle at the time of writing, so this section carries the method and no results. It runs before results appear here, and the wrong answers stay in when they do. A panel with no errors in it is an advertisement, not a measurement.

Where does this approach fail, and who should skip it?

Entity correction fails in three situations, and it is worth knowing which you are in before anyone scopes the work.

It fails when the wrong description is actually right. If assistants describe you as a generic version of your category, the likeliest explanation is that your public material is one. Schema does not fix that.

It fails against volume. A company with a decade of press coverage cannot outweigh its own history with a canonical record and four profile edits. You can shift the margin, not rewrite the corpus.

It fails against a name collision with a bigger entity. If a larger company or a common phrase owns your string, markup will not win it back.

The honest ceiling: this is a hygiene practice. It makes the descriptions you can influence accurate and consistent, which protects deals that would otherwise die on one wrong sentence. It does not make an assistant recommend you for a commercial query, because those citations mostly go to third parties.

How long does correction take, and what will not change?

Unknown, and be suspicious of precision. No vendor publishes re-crawl frequency for OAI-SearchBot, Claude-SearchBot or PerplexityBot, model updates confound every before and after, and the same page can be quoted by one assistant while another has never heard of it.

What is knowable is how to remove discovery delay on the Bing dependent surfaces, which feed Copilot and ChatGPT's candidate retrieval. Submit sitemaps to Bing Webmaster Tools, use IndexNow on publish and update, and switch on the AI Performance report there plus the generative AI report in Search Console, because neither works retroactively.