An entity in AI search is a specific thing in the world that a machine can distinguish from every other thing with a similar name, and your brand is not automatically one. Until a retrieval system can tell that your company is a distinct organisation rather than a word that happens to appear on some pages, everything downstream of that is unreliable, including whether an answer about your category can safely name you at all.
The distinction worth holding onto is between a string and an entity. A string is a sequence of characters. An entity is a thing with an identity, properties and relationships. Your brand name starts life as a string, and something has to happen for it to become the other kind of object. Most B2B software companies have never done that something, which is why this is one of the few remaining pieces of AEO with almost no competition.
Three steps your name has to survive
This is not a marketing framework. It is how the field of entity linking has been described in the research literature for over a decade, and the survey work in this area breaks the task into three subtasks that run in order. Getting them in the right order explains a lot of otherwise baffling behaviour.
First, recognition. Something has to notice that a run of characters is the name of a thing rather than ordinary prose. Second, candidate generation. The system pulls a shortlist of known entities that name might refer to. Third, disambiguation, where it picks which candidate is the right one, using surrounding context to choose. The standard survey treatment of entity linking lays this out in detail, and the practical point is that a brand can fail at any of the three for completely different reasons.
| Step | What has to happen | How a brand fails it | What fixes it |
|---|---|---|---|
| Recognition | Your name is identified as naming a thing | The name is an ordinary word, so it reads as prose | Consistent capitalisation and a stated legal name |
| Candidate generation | Your company is on the shortlist of possible matches | You are in no reference record anywhere | Presence in registries and third-party databases |
| Disambiguation | You are chosen over the other candidates | A larger organisation shares your name | Explicit identity links and consistent context |
The five-minute test. Ask any assistant what your company is, using only the name and no other context. If the answer is blank, hedged, or about somebody else, you have an entity problem and no amount of content work will address it. If it is accurate and specific, this article is not your bottleneck.
Why software brands fail this more than most
Four naming patterns cause almost all of it, and the software industry has a weakness for every one.
Names that are ordinary words. Notion, Slack, Stripe and Monday all had this problem early and solved it with sheer volume of coverage. A company at an earlier stage does not have that option, because recognition fails before anything else gets a chance.
Names shared with a larger organisation in another sector. Disambiguation resolves toward whichever entity has more supporting evidence, which will not be you. Every answer about you competes with an unrelated company that is better documented.
Recent renames and rebrands. The old name has years of accumulated evidence attached to it and the new one has almost none, so systems keep resolving to the previous identity or treat the two as unrelated things. This is the most avoidable version of the problem and the one most often walked into.
Product names that differ from the company name. Buyers ask about one and your material talks about the other, and nothing states that the two are connected. That connection is trivially expressible in structured data and almost never expressed.
If one of those four patterns describes your company, you already know which one, and the useful next question is how badly it is currently costing you. That is the entity half of a free visibility check.
What actually resolves an entity
The mechanism is unglamorous. You are trying to make it cheap for a machine to be confident, which means stating the same facts in the same way everywhere and pointing at records it already trusts.
The move with the largest effect is an explicit identity link. Structured data supports saying this name refers to the same thing as that record, and the record most widely read for this purpose is Wikidata, which assigns a stable identifier to each entity and is consumed by systems that need to resolve names to things. You cannot simply create a record for yourself, since inclusion has independent notability rules, and knowing whether one exists is a check most companies have never run.
Beyond that, identity has a more formal layer that rarely gets mentioned in marketing conversations. Registry-grade identifiers exist for legal entities, the clearest example being the Legal Entity Identifier, a twenty-character code that uniquely identifies a legally distinct organisation. Answer engines are not primarily reading LEI records, so treat this as context rather than a tactic. The reason it is worth knowing is that it shows what unambiguous identity looks like when a system genuinely requires it, and it is a useful standard to hold your own entity data against.
The rest is consistency, which sounds trivial and is where most of the failure actually is. One legal name, used the same way everywhere. One founding date. One canonical URL that never moves. The same description on your own site as in every profile you control. Where a third-party record contradicts you, fix the record rather than adding a fifth version of the truth.
How wide open this actually is
Two numbers make the case better than any argument. Crawl data from the HTTP Archive Web Almanac for 2024 found structured data in JSON-LD form on 41 percent of pages, up from 34 percent two years earlier. Pages using that structured data to point at a Wikidata record for their entity sat at 0.17 percent, and at Wikipedia 0.13 percent.
So two pages in five carry machine-readable markup, and fewer than two in a thousand use it to say which thing in the world they are. The industry adopted the syntax enthusiastically and skipped the part that resolves identity, which is the part that matters when an answer has to name somebody.
That gap is not there because the work is hard. It is there because it is invisible in every dashboard. Nobody gets promoted for adding four lines to a JSON block, and the effect does not show up as a traffic spike the following week, so it loses every prioritisation argument it enters.
A week of work, in order
- Run the recognition test. Ask four assistants what your company is, with no other context. Write down each answer verbatim, because the differences between engines tell you which step is failing.
- Search your own name and see what else answers to it. If a bank, a band or a bigger firm in another sector shares it, you have a disambiguation problem and you now know the specific competitor for your own identity.
- Check whether a reference record exists. Look yourself up in the structured knowledge bases. Absent is common and is itself the finding, since candidate generation cannot shortlist a company it has no record of.
- Write one canonical set of facts. Legal name, trading name, founding year, headquarters, what you sell and to whom. One version, in one place, that everything else copies.
- Put it in structured data with identity links. A single Organization block with a stable identifier and a filled-in list of records that refer to the same entity.
- Reconcile the profiles you already control. Every third-party listing repeating an old description is evidence pointing the wrong way, and each one you correct is a vote for the right answer.
Step four is the one teams underestimate. Getting six people to agree on a single sentence describing what the company does is harder than any of the technical work here, and until that sentence exists there is nothing consistent to publish.
Step one takes five minutes
Ask four assistants what your company is and you will know almost immediately whether you have an entity problem. If the answers come back blank, hedged or about somebody else, our free visibility check includes the entity diagnosis and identifies which of the three steps your name is failing at, which determines whether this is a week of work or a naming conversation.
Why this is worth doing before content work
Because entity resolution sits underneath everything else rather than beside it. A perfectly extractable passage attached to an ambiguous identity can still produce an answer that credits the wrong company, and that outcome is worse than not being cited, since your work is now supporting somebody else’s name.
It is also close to a one-off task. Crawler policy needs revisiting as providers ship new agents and extractability is ongoing editorial discipline, whereas entity resolution is mostly a week of unglamorous work that then holds for years. Cheap, durable and almost universally skipped, which is a rare combination in this field.
The honest caveat is that it will not on its own get you recommended. Being a well-resolved entity means a system can talk about you accurately and confidently. Whether it chooses to put you forward as the answer depends on what other people have written, and that is a separate and slower problem.
What it does buy you is the right to compete. An ambiguous company is not losing to better competitors, it is failing to be considered, and those are different problems that feel identical from the inside. Resolve the identity and you at least find out which one you actually have.
Entity work is the least visible thing on this site and the closest to a one-off. Do it once, keep it consistent, and it stops being a problem for years.
We publish our entity audits as we run them, including the ones where the answer was that the client shared a name with a bank. The newsletter is where they go out.
Related reading
- AEO vs SEO. Why the unit changed from a page to a passage, and the prize from a click to a mention.
Shaban runs answer engine optimisation for B2B SaaS companies at Uncited Brands. He works the unglamorous end of the problem, mostly crawler access, retrieval diagnostics and measurement baselines, and publishes the tests that do not go his way alongside the ones that do.