The thesis
Machines answer about the entity behind your name.
Entity SEO is the work of making machines agree on who you are. Before an answer engine writes a sentence about your company, it resolves your name to an entity: a category, a location, a set of people and facts. If that resolution is wrong, the answer is wrong, and no amount of good copy fixes it.
Your brand is a string. The entity behind it is what gets answered.
01
Entity SEO starts with the identity a machine resolves
An answer engine does not read your homepage and then describe you. It decides which "you" it is describing, then goes looking for facts that fit. That first step runs before any of your positioning gets a vote.
Google says this out loud in its own documentation. The guidance for Organization structured data states that adding it to your home page "can help Google better understand your organization's administrative details and disambiguate your organization in search results." It goes further, noting that certain properties exist purely "to disambiguate your organization from other organizations" (Google Search Central). Read that as a product requirement. Disambiguation is a job the system performs whether or not you help, and Organization markup is the interface it exposes for helping.
Most teams treat identity as something implied by good design. Identity is an input you declare: legal name, alternate names, category, address, contact, the profiles that belong to you. Declare it once, in machine-readable form, on the page a crawler reaches first. Then go check that the answer engines you care about are quoting the right company.
Nobody writes a brief for the question "which company is this?" Every answer engine answers it first.
Try thisAsk three different AI assistants to describe your company in two sentences. If any of them describes a business you do not run, you have an entity problem sitting upstream of every content problem on your list.
02
The knowledge graph is a record you feed but do not own
Google's Knowledge Graph is a database with a row for you in it, and you are not the administrator. By Google's own published figure it "has amassed over 500 billion facts about five billion entities," assembled from material shared across the web plus open source and licensed databases (Google, 2020).
Two things follow. The graph can be wrong about you, and Google says so plainly in the same post: inaccuracies can happen, and the company accepts reports about them. There is also a formal path back. Many knowledge panels can be claimed by the subject they describe, which hands you a channel to correct facts and suggest changes yourself. Very few mid-market teams have walked that path. It costs almost nothing, and the panel is often the first thing a buyer reads about you.
The habit I argued for in the first column of this series applies to your own record. Take nobody's word for it, the graph's included. Go look at what it says about you.
There is a row about your company in a database you cannot log into. Somebody on your team should read it.
Try thisSearch your company name in an incognito window and read the knowledge panel line by line. Note every fact that is stale, missing, or belongs to someone else, then check whether the panel has been claimed.
03
Ambiguous names get a default reading, and it may not be yours
If your name collides with anything, a common word, a surname, a bigger company in another sector, you are competing with a default.
Researchers at the University of Vienna and LMU Munich tested this directly. Their evaluation protocol separates whether a model knows a fact from whether it applies that fact. Run against 49 ambiguous entities, state-of-the-art LLMs "perform poorly with ambiguous prompts, achieving only 80% accuracy," with "significant biases for preferred readings" and self-inconsistency across their own answers (Sedova et al., arXiv, 2024). The finding that should worry a marketer is where the preference comes from: the preferred reading tracks how often that meaning shows up in the training data.
So the model has a favorite interpretation of your name, earned by frequency. There is no arguing with frequency. What you can change is availability: the category stated plainly, the disambiguating facts attached, the name used consistently in the places that get ingested. Unglamorous work, and a decent share of what the teams in Innovative Group's digital marketing and technology practice end up untangling on a new account.
A model does not weigh your name. It picks the reading it has seen most and moves on.
Try thisList every other thing your company name could reasonably refer to. That list is your real competitive set for the disambiguation step, and it has almost nothing to do with your actual competitors.
04
Entity drift is a defect class, and Google says so in writing
Here is where the engineer in me gets loud. Entity facts drift the way a footer drifts: quietly, one edit at a time, until five surfaces disagree. A title changes on LinkedIn but not in the team page schema. An office moves and the old address survives on a directory listing. A service line gets renamed in the nav and nowhere else. Each one is a defect with a slow bill attached.
Google treats it as a defect too. The general structured data guidelines are explicit: "your structured data must be a true representation of the page content." They tell you plainly not to mark up content invisible to readers, using the example that if the JSON-LD describes a performer, the HTML body must describe that same performer. They ask you to "provide up-to-date information." And they carry teeth. A structured data issue can trigger a manual action that strips the page's eligibility for rich results (Google Search Central). Markup that quietly stopped matching the page is a validation failure with a penalty attached.
The fix is a defect log. Write the canonical facts down once, list every surface that repeats them, and put a recurring check on the calendar. That is the systemizing argument from an earlier column pointed at identity, and it is exactly the kind of recurring check a Next Best Action program is built to carry so it stops depending on somebody remembering.
Nobody schedules a review of what your company officially claims to be. That is precisely why it drifts.
Try thisPick five facts (legal name, category, address, phone, leadership titles) and check them across your homepage schema, your team page, LinkedIn, your Google panel, and one directory. Count the disagreements. That number is your entity defect count.
05
Conflicting facts about the same name break the answer outright
Inconsistency does more than blur the answer. It can collapse it.
A team at VinAI Research and Nanyang Technological University built a benchmark, WhoQA, that induces exactly this failure. It asks about a property shared by several real entities that happen to have the same name, then measures what the model does with the conflicting sources. The drop is severe. One seven-billion-parameter model answered 91.8 percent of the questions correctly when a single clean context was supplied, and 12.7 percent once the conflicting contexts were added without warning. Larger models held up better and still lost ground. The paper also notes that some models simply pick one answer among many and present it, which the authors flag as more concerning than declining to answer, because it produces confident misinformation (Pham et al., arXiv, 2024).
That is a lab setting with same-name people, and the mechanism transfers cleanly. When two of your own surfaces state different founding years, different headquarters, or different service categories, you have manufactured a knowledge conflict about yourself. A retrieval system pulling both will hedge, pick, or hand the question to a company whose facts agree with each other. This is also the quiet risk in a fractional leadership handoff: each new team adds a surface, and reconciling the old ones rarely has an owner.
Two versions of your address is not a rounding error. It is a coin flip you did not know you were running.
Try thisPut your five canonical facts in one column and every place they appear in another. Anywhere the row disagrees with itself, you are generating the conflict, and you are the only one who can retire it.
06
Category association decides which questions you appear in
The freshest evidence for all of this landed a few days ago, published as a visibility story. Underneath, it is a story about classification.
A study run by the agency Fractl and published in Search Engine Land on August 17, 2026 tested how consistently models recommend brands. The method: 96 industry-specific prompts across eight industries, each run 15 times against GPT-4o, Gemini 2.5 Flash, and Claude Sonnet 4.6, producing 4,320 responses and more than 8,500 unique brand references, then mapped against Ahrefs authority data. Roughly nine in ten brands behaved as expected, stronger search authority tracking stronger AI visibility. The edges are where it gets interesting. About 471 brands were badly underexposed despite domain ratings above 80 and millions of monthly organic visits, while 377 punched far above their traditional signals. Microsoft and Spotify topped the underrepresented list for FinTech prompts, and the author's diagnosis is precise: the models had decided what counts as a fintech, and those names were not on the list (Libert, Search Engine Land, 2026). Worth naming the provenance: this is a single-agency dataset, and the author discloses that she co-founded the agency that ran it.
Treat that finding as a content shortfall and you go off to write more. Treat it as an identity failure and the work changes. Category is a property of the entity, and the entity is what the model consults before it builds a shortlist. Being absent from the queries your buyers actually ask is frequently a categorization failure, and more content filed under the wrong category will not move it. The remedy is the same discipline this column keeps landing on, now pointed at classification: state the category explicitly, attach it to the entity, and make the association repeat in the places models learn from. The argument I made in an earlier column about building authority beyond your own walls is the distribution half of this. Category correctness is the identity half. If nobody has checked which category the models have filed you under, that is a good first conversation to have with our team.
You can win every keyword in your category and still not be in it, as far as the model is concerned.
Try thisAsk each major assistant for the top five companies in your category, then ask it to describe your company. If you are absent from the first answer and correctly described in the second, you are filed somewhere else.
Ashlesha Khond, Innovative Group