Entity or non-existence: a 7-step diagnostic for whether you exist to the machine

An analyst in Zurich opens his laptop at eleven in the evening. Cold pizza on the desk, the phone beside it buzzing with a message from a colleague: “Check what ChatGPT says about us before tomorrow’s investor call.” He types his company’s name. The machine responds in an even, confident voice — three paragraphs about a completely different firm with the same word in its name, registered in Delaware, with which his team has nothing in common. He clears it, rephrases the question. The machine adds an invented founder. The analyst stares at the screen the way you stare at a letter with someone else’s name in your own mailbox.

Meanwhile, deep in the index feeding that answer, there are over 120 million entities. That is exactly how many objects — people, companies, places, concepts — Wikidata, the world’s largest open knowledge graph, contains, with approximately 1.65 billion statements about the relationships between them (Wikidata:Statistics). The Zurich analyst’s company is not among them. To the machine it is not wrong. It simply does not exist.

It is tempting to blame bad SEO or the “algorithm not liking us” — and to get the address completely wrong. What is at stake is the entity recognition mechanism: how a system decides whether you are a distinct, identifiable object in the world, or merely a random string of characters. This is exactly what we habitually disguise with the alibi “we just haven’t been indexed yet.” I will unpack it through three lenses: engineering (how the graph is actually assembled), reputational (what an investor, client, or journalist who Googles you sees), and temporal (why the action window is narrow right now). The thesis-blade: invisibility to the machine is neither a verdict nor a mystery — it can be diagnosed in one evening, the gaps are usually exactly three, and the first week of targeted action is worth more than a year of expensive SEO.

The scene doesn’t lie — your intuition about it does

The first temptation is to explain the void by scale. “We’re a small company — why would we appear in the graph?” But scale is irrelevant. The knowledge graph contains rural libraries, individual beetle species, district councillors from three centuries ago. The entry threshold is not fame but distinguishability: is there enough consistent, structured information for the system to confidently say “this is a distinct object, not a variant of something else”?

The second temptation is worse — explaining the void with time. “We just haven’t been indexed yet.” This is an alibi that costs companies years. Because indexing a page and recognising an entity are two different machines. Google sees your page within a day. But the decision “this is Dnister, a distinct company, not that other firm with the same word” is made separately, based on signals you most likely never consciously provided. Time alone does not resolve this. You can write an honest blog for ten years and remain nothing but fog to the machine.

Let’s call the mechanism by its name. The industry calls it entity resolution — resolving an entity. It is the process in which the system takes all mentions of a word that looks like your name and tries to merge them into one object or, conversely, split them into several. If signals are consistent — one clear entity is born. If they contradict each other — the system either creates three different objects or creates none. The Zurich analyst saw the third scenario: the machine could not find enough consistency to construct his company, and defaulted to the nearest confident object — someone else’s.

There is an important nuance that people miss: the machine works not with your intentions but with your matches. It does not ask what you meant when you named the company. It counts how many independent sources repeat the same combination of facts — name, domain, founding year, city, CEO’s name — and how consistently those combinations agree with each other. One strong match is worth more than a hundred weak mentions. That is why a company with thirty press articles but not a single consistent record in an authoritative graph can lose the recognition game to a small startup that carefully filled in five fields across five databases, and all five agreed.

How it actually works: three layers, not one Google

The common mental model: “there’s Google, it knows everything, you need to please it.” Reality is three separate layers, and they are not synchronised.

The first layer — Google’s Knowledge Graph. This is its internal entity database, the source of those panels on the right side of search. Google feeds it from many sources, but one of the most powerful is Wikidata: the more external profiles consistently point to the same company, the higher the entity’s “confidence score” in the graph (Stackmatix, Organization Schema & Knowledge Graph 2026).

The second layer — large language models like ChatGPT. And here is the counterintuitive part: they do not query Wikidata or Google’s graph live. According to available tests, OpenAI rather takes a “snapshot” of these sources at the time of model training, adds its own collected data — and freezes. The check is simple: there are objects present in both Wikidata and Google’s graph that ChatGPT still does not recognise as entities, because they did not exist at the time of the snapshot (llmrefs, ChatGPT Entities). This means that between your appearance in the graph and your appearance “in the model’s memory” lies a lag equal in length to one training cycle — months.

The third layer — AI-generated answers in search itself (AI Overviews, AI Mode). Here the pace is the opposite: explosive. Measurements differ by methodology — Semrush sees a peak of around 25% of queries in summer 2025, while the independent tracker Advanced Web Ranking shows almost half of all results by year end (Search Engine Land, AI Overviews 2025 data). They agree on one point: within a year, AI overviews went from a rare exception to a mass surface on which the machine speaks about you, not your site.

Combine the three layers and you get a trap. The surface on which AI describes you is growing fast. The source of truth for those descriptions (the graph) updates slowly. And the models’ “memory” freezes for months. If you are not an entity, this machinery does not stay silent about you — it improvises confidently. Exactly as it did in Zurich.

The evidence staircase: from “this is theory” to “this is already costing money”

One could object: “fine, the graph has its own life, but there are no real losses from this.” Let’s climb the evidence staircase, from the abstract to what hits revenue.

First step, behavioural. Search stops being a door to your site. According to Semrush’s 2025 research, 58.5% of search sessions in the US and 59.7% in the EU end directly on the results page, with no click through (Semrush zero-click, via Click-Vision). The user read the machine’s answer — and left. If you are not in that answer, you might as well not have existed.

Second step, causal. Similarweb recorded that since AI Overviews launched, the share of searches with no click at all rose from 56% to 69% in just one year (Similarweb, via Stan Ventures), and in Google’s new AI Mode it reaches 93% (Semrush 2025, via Click-Vision). Meaning: where the machine speaks, people almost never go to check the primary source.

Third step, monetary. On queries with an AI overview, organic CTR collapsed by approximately 61% (Ekamoira, Zero-Click 2026). This is no longer about abstract “visibility” — it is about two-thirds of the traffic that used to reach you, now settling inside someone else’s paraphrase.

A safeguard is needed here, because evidence staircases like to collapse into elegant lies. The growth of AI overviews correlates with the drop in clicks — but this does not prove that your specific invisibility in the graph cost you a particular client. Perhaps that client would not have come anyway. Industry-level correlation is not causation at the level of your company. What the numbers do prove with certainty is something else: an environment in which decisions about you are increasingly made before anyone opens your site. And if you do not exist as an entity in that environment — the decision will be made without you, and, as in Zurich, possibly about someone else.

The mirror: recognise your own evening

Now let’s come down from the industry level to the level of your laptop. Do this tonight — it will take less time than an episode of a TV series.

Open ChatGPT and ask simply: “What is [your company]?” Look not at the quality of the answer, but at one thing — is the machine confident, and is it talking about you? Three typical results, each a diagnosis.

If the machine confidently describes someone else — you have an entity collision. Your name shares space with a stronger same-named object, and the system consistently picks that one. This is the most insidious gap: you feel like you “exist on the internet” because an answer comes back. It’s just that the answer isn’t about you.

If the machine hesitates, mixes up facts, stitches you together with adjacent companies — you have a blurred entity. Signals about you exist, but they contradict each other: on your site you’re “Dnister Strategy,” on LinkedIn you’re “Dnister,” in the directory you’re “LLC Dnister,” and the address differs across three places. The system cannot merge this into one object and produces fog. This is the classic disease of inconsistent name, address, and contact information — the same signal conflict that reduces a graph’s confidence (digitalapplied, Entity SEO 2026). It is worth knowing which signal the machine reads most carefully: the sameAs property in structured markup — a direct declaration that “this site, this Wikidata record, this LinkedIn profile, and this registration entry are one and the same object.” This is the strongest available identity signal: it lets the system triangulate you across several trusted sources and build confidence from exactly that (Stackmatix, Organization Schema 2026). When sameAs contradicts itself or leads nowhere — triangulation fails and fog is born.

If the machine honestly says “I have no data” or invents from scratch — you have an empty entity. You appear in no authoritative graph; no consistent external profile points to you. Paradoxically, this is the cleanest of the three diagnoses: treating emptiness is easier than untangling a collision. A clean slate is simpler to fill than convincing the machine that a strong same-named neighbour is not you.

A small but important trap at this stage: don’t confuse the machine’s confident tone with the truthfulness of its content. A large language model is designed to sound equally calm whether it’s citing a verified fact or inventing a non-existent founder. Tone is not a truth signal. That is why diagnosis must come not from the impression “sounds convincing,” but from cross-checking: open two or three sources alongside it and verify whether the machine is paraphrasing real facts or stitching together plausible fiction. In exactly that gap — between the confidence of form and the emptiness of content — most reputational surprises live.

Two minutes in ChatGPT and you already have your first hypothesis. Now the full pass: seven steps, one evening.

#	Diagnostic step	Where to check	What “bad” looks like
1	Direct question to the model	ChatGPT: “What is [company]?”	Someone else’s company / invented facts / “I don’t know”
2	Comparison question	ChatGPT: “[you] vs [competitor]“	You’re absent from the comparison or confused with someone else
3	Google Knowledge Graph	Search the name + check for right-hand panel	No panel, or panel contains someone else’s facts
4	Wikidata	wikidata.org, search the name	No entry, or entry is empty/outdated
5	Name and address consistency	Site / LinkedIn / directories side by side	Name and address differ across sources
6	External sameAs links	Do profiles point to each other	Profiles are isolated, not cross-linked
7	Site structured markup	View source of the homepage	No Organization markup, or it’s incorrect

Seven fields. At the end you have not a feeling but a map: exactly where the chain breaks between “you exist” and “the machine knows it.” In the vast majority of cases the break is in exactly three places — collision, blur, emptiness — not fifteen, as agencies like to warn.

The antagonist without a face — and why you can’t negotiate with it

This story has an antagonist, but it is not malicious. It is the entity recognition system — cold, consistent, without ill intent. It did not decide to ignore you. It simply did not receive enough consistent signals to risk calling you a distinct object — and defaulted to the conservative option: either picked a more confident neighbour, or said nothing.

You cannot negotiate with this antagonist using words, complaints, or outrage. It does not read your press releases about how unique you are. It reads only structure: does your name match across ten sources, do those sources point to each other, is there a graph entry to lean on. The machine does not care that you know who you are. It needs independent sources to say it consistently.

Hence the dark comedy of this situation: you can work honestly for twenty years, have real clients, an office, tax returns — and remain less real to the machine than a beetle that one PhD student carefully entered into Wikidata one evening. The reality of your business and your reality to the machine are two different kinds of existence, and the second is granted to no one by virtue of seniority.

Why right now: the 2018–2026 window

Why couldn’t this article have been written in 2018 with the same urgency? Because until recently, invisibility to the graph was cosmetic. So no panel — the person still clicked through to your site from the results and read your version of yourself.

The break came between 2023 and 2026. In one year the share of searches without a click rose from 56% to 69% (Similarweb, via Stan Ventures) — meaning the machine’s intermediate summary stopped being an exception and became the norm for “how a person learns about you.” Previously, the knowledge graph was a showcase. Now it is the script by which the machine speaks about you aloud to a third party, while you have no idea.

The second window is model lag. Since ChatGPT and others work from a “snapshot” of the graph at training time (llmrefs), everything you fix today will enter the “awareness” of the next generation of models only months from now. Every week of delay is not a lost week — it is a lost week multiplied by the model’s training cycle. That is exactly why the window is narrow right now: you earn the right to appear in tomorrow’s model through tonight’s actions.

Re-plating: the first week is worth more than a year of SEO

Let’s put everything back on the plate. The diagnosis — one evening. The treatment — also not a marathon, despite what is sold as a “year-long SEO strategy.”

Here is the timing mechanism, and it is precise. When you make consistency edits — fixing Wikidata, aligning name and address across sources, adding Organization markup, cross-linking profiles with mutual references — the machine responds in steps: re-crawling of pages starts within 1–2 days, entity processing within 3–7 days, and the first movements in the Knowledge Panel become visible within 2–4 weeks (instantpress, Knowledge Panel process 2026). Full panel appearance for a company is more of a 6–12 month horizon, and for a person 12–24, since the recognisability threshold for individuals is higher (instantpress). One separate observation: consistent sameAs links across 8–10 authoritative databases produced measurable panel movement within 45 days (digitalapplied). ⚠VERIFY: the specific timelines in all these sources are practitioner observations, not a guaranteed Google SLA; “no deterministic schedule exists,” as the authors themselves acknowledge.

From this comes the thesis worth putting above your desk: the first week of targeted work on entity structure produces more movement than a year of expensive content SEO on a company the machine still cannot distinguish from a namesake. You cannot optimise the ranking of an object that doesn’t exist for the system — you must first earn the right to be a distinct entity, and only then does fighting for positions make sense. A year of blog posts for a fog-company is polishing a mirror that has no reflection.

This is the right moment to note that exactly at this intersection — between “a business exists” and “an entity exists for the machine” — a distinct discipline has emerged. When I was working through the logic of entity audits with the WikiBusiness team, their most common finding was not “your content is weak” but “your company is splitting into three different objects across three different graphs” — the same collision diagnosis we started with. Symptomatically, businesses arrive asking “improve our rankings,” and it turns out there is nothing to improve: the machine has not stitched them into a single whole. First — stitch. Only then — talk about positions.

What would prove this wrong

An honest author leaves the door open to refutation. What would break this thesis?

If models switched to live graph queries instead of snapshots — the lag would disappear, and “the first week” would lose its weight: fix it today, the machine sees it tomorrow. If Google began reliably recognising an entity from a single authoritative source without requiring multi-source consistency — the “blurred entity” diagnosis would lose its meaning. And if zero-click search rolled back to pre-AI-overview levels — the entire urgency of “why right now” would deflate, and you could go back to not worrying about this for years.

None of these three scenarios has materialised; the first and second are technically possible, the third is not. Until they do — a one-evening diagnostic remains the week’s best investment.

Hard kicker

The Zurich analyst will close his laptop around one in the morning. Not because the situation is hopeless — but because in those two hours he understood exactly three things: where the chain breaks, that it breaks in three familiar places, and that tomorrow’s investor call will happen before the machine has any chance of learning his company. He will write one line to his colleague and turn off the light.

And that other company from Delaware will go on sleeping peacefully in the knowledge graph — confident, distinct, with its own panel — not suspecting for a moment that for half an hour tonight it was someone else entirely.

---

Partnership disclosure: WikiBusiness is a recurring partner of NeuroDrift. The mentions in this text are used as illustrations of entity audit mechanics, not as a paid recommendation.