Knowledge Graph ve Varlık Yönetimi: AI Sizi Tanıyor mu

Screen share with the owner of a dental clinic. He types his clinic's name into ChatGPT: "Tell me about Ahmet Diş Kliniği." The answer is polite and completely empty — several places may share that name, which city did we mean? Same name on Google: no knowledge panel on the right. Yet there is a Maps listing, a site, a three-year blog archive, and on some searches they rank first.

What is missing is not content but identity. To search and answer systems, "Ahmet Diş Kliniği" is not yet a business in Kadıköy with a known founder and address; it is a three-word string. That distinction is entity recognition, not keyword ranking.

This article covers why the gap forms and which tools close it. A warning up front: none of it guarantees a knowledge panel. No method does; there are only things that raise the chance of recognition and things that prevent it.

What an entity is, and how it differs

A keyword is a string: "kadıköy implant" either appears in a text or not; the match is at character level.

An entity is a node: something existing in the world, with properties of its own and relations to other nodes. A business, a person, a place, a product. Recognised as an entity, "Ahmet Diş Kliniği" holds a record roughly like this: a dental clinic, in Kadıköy, Istanbul, founded by this person on this date, official site here.

The difference shows up here. A language model reading your page sees your brand name, but decides which entity to attach it to independently of your page. Finding no node, it either drops the name or hedges: "there may be more than one business with this name." Both give the same outcome — the first condition for your brand being recommended in ChatGPT and Gemini is unmet.

The reverse happens too: first in the organic list for "kadıköy dental implant", unmentioned in an AI answer the same week. Ranking is a page-level contest; entity recognition is a brand-level record. They are won separately.

How a Google Knowledge Panel forms, and why most SMEs lack one

That box to the right of the search results is not built from your site. It comes from Google's own Knowledge Graph — what Google is confident enough about for an entity.

Two boxes get confused:

• Local business card: from your Google Maps / Business Profile listing. Address, opening hours, reviews, phone. You manage it directly; most small businesses have one.
• Knowledge panel: about the entity itself. Founding date, founder, parent company, related entities. You cannot create it; Google does.

It fails to form for most small businesses not through penalty but because you cannot clear the bar: before opening an entity record Google generally waits for more than one independent source to say the same thing. A business whose only source is its own site stays under it — self-description is a claim, not a verification.

The honest sentence: nobody can undertake that a knowledge panel will form. An offer promising "do these steps and your panel arrives" sells an output it does not control. You can make the signals beneath it clean and consistent; the rest is not yours. If one forms, the "claim this knowledge panel" link lets you verify and edit some fields. Before that there is nothing to claim.

Wikidata: what it is good for, when it backfires

Wikidata is a machine-readable, openly licensed, structured knowledge database. Each record carries a Q number and consists of statements: instance of → dental clinic, country → Turkey, official website → …, inception → ….

Its weight comes from many systems using it as a clean source for resolving entities. A record can anchor scattered mentions to one identity.

But there is a condition: notability. Wikidata's rule looks for one of three cases — the subject appears in another Wikimedia project, or it is a clearly identifiable entity verifiable against a serious and publicly available source, or it matches a valid identifier in a structured external database. A record opened for a business no independent source mentions ends in a deletion discussion, usually deleted.

Faking a record is a bad tactic, because:

• On Wikidata every edit's history and author are public; who added what is visible.
• The source link you give gets clicked and checked. A fabricated reference is exposed in minutes.
• A deleted record never returns; the deletion record stays. The effort turns negative.

The right order is the reverse: source first, record after. With the source in hand, fill in the official site, headquarters location, founding date, field of activity and any external identifiers — each sourced. And expect this: a Wikidata record does not produce a knowledge panel by itself. One brick, not the building.

The Organization schema and sameAs: merging scattered identities

The most direct work you can do on your own site. The Organization schema makes your "this is who I am" statement machine-readable; sameAs says "these profiles are also me." Scattered mentions collect into one entity.

The fields to fill:

Field	What goes in it
name	Your everyday brand name, in one spelling
legalName	Full trade-register title, including A.Ş. / Ltd. Şti. (Turkish for Inc. / Ltd.)
url	Canonical home page address
logo	Directly reachable, permanent image address
sameAs	Social profiles, directory listings, professional chamber page, press mentions
address	Open address as PostalAddress — district, province, postcode, country
founder	Founder(s), with the Person type
foundingDate	ISO date (2019-03-01)

Three rules for sameAs: list only pages that genuinely belong to you or mention you; keep them alive (a dead account is noise, not signal); use exactly the same name on all. Define the schema with one canonical @id and reuse it site-wide — a different Organization block per page paints several entities, not one. For schema mistakes and the E-E-A-T link, see the role schema markup plays in AI engines.

Name and NAP consistency: how an entity splits in two

Five spellings of one business in five places produce not one entity but five weak ones. In Turkish the risk is higher: the capital İ/I and accented letters break the string easily.

An audit list — compare each line across five sources (site, Google Business Profile, social media, directories, invoices):

1. Brand name spelling: "NextGeo", "Next Geo" or "NEXTGEO"? Pick one.
2. Turkish characters: "Işık" and "İşık" are different strings. Whichever is correct, use it everywhere.
3. Legal suffix: "… Ltd. Şti." appears in some places, not others. Use the legalName / name distinction; do not mix them.
4. Address format: order of neighbourhood/street/number, abbreviations ("Cd." – "Caddesi", Turkish for avenue), building name. Pick one.
5. Phone format: keep the international +90 216 … form everywhere.
6. Old address and old branch: if you moved, correcting the old listing matters more than adding the new one.
7. Domain: www or no www, http or https — one canonical form.

The boring but highest-return part of the job. Aligning existing records, without producing new content, can visibly improve an entity's resolution.

Independent verification: if your site is the only source

Everything you write on your own site is a claim. Finding no second voice to confirm it, the system plays safe — retreating to a safer answer instead of recommending you.

Sources with verification value:

• Official and professional registers: trade register records, MERSİS (Turkey's central business registry), bar association / medical chamber / chamber of commerce member pages, licence and authorisation lists.
• Publications with an editorial filter: local and sector news sites, trade magazines with an editor.
• Institutional third parties: university, municipality and chamber event pages; conference speaker lists; interview and podcast episode pages.
• User-generated but moderated: Google reviews belong here; volume and content both carry a signal. Detail in how Google reviews affect AI recommendations.

Low weight: paid listing directories, dozens of sites copying the same press release, profile pages you filled in yourself. These produce repetition, not verification.

A fixed number would be wrong — no rule says "five sources are enough." In practice two independent, serious sources beat twenty weak links copying each other.

Confused with the wrong entity — how to fix it

Three typical scenarios:

Another business with the same name. The model has merged two records and gives a service or city that is not yours. Sharpen the distinguishing information: fill address, founder and foundingDate in the schema; put the founding year, district and founder's name in the about page in plain sentences; always name the district beside the brand in your content.

The wrong city. Usually an old directory listing or a news item from before a move. Correct the directories one by one; the output does not change until the source does.

A closed branch. Do not delete the Google Business Profile listing, mark it permanently closed. A deleted record hangs in a void; one marked closed carries information.

Feedback channels are limited but real: Google Business Profile edits, the feedback link in knowledge panels, a sourced Wikidata edit, a Bing Places listing. And plainly: you cannot edit an AI answer directly. You can only change the sources feeding it, then wait for a recrawl.

How long it takes, how it is measured

Be suspicious of anyone giving an exact figure on timing. The rough observed band: schema and directory corrections get crawled within weeks; entity trust moves on a scale of months; old information inside a model's training cut can stay far longer. It varies by business and by the sector's source density.

For measurement, set a fixed prompt set — 15-20 questions, once a month, on the same engines. Record four things each run: (1) did the brand name come up, (2) are the city and service correct, (3) was it confused with another business, (4) was your own domain cited. Add brand-name query impressions from Search Console and your direct traffic.

And something we say not to measure: a single chat answer. Ask the same question twice and you can get two different answers; a trend from one sample misleads. Reporting that ignores this shows undone work as done — one way to tell whether your current agency is really doing GEO.

Entity work comes before content production. Twenty articles from an unrecognised brand are twenty texts with no owner. Content produced after the identity is cleaned and sources aligned attaches to something. We run that order at enterprise level; see the solutions page.

Frequently Asked Questions

Is the Knowledge Graph the same as keyword ranking?

No, they work on separate layers. Keyword ranking is a page-level contest: which URL fits a given query best. The Knowledge Graph is a brand-level identity record; it decides whether your brand is recognised as a business existing in the world. You can rank first for a query and go unmentioned in AI answers the same week, because the two are won separately.

How do I get a Google knowledge panel?

You cannot create the panel; Google creates it once enough confidence has formed about an entity, and no method guarantees that. You can fix the signals it rests on: a complete Organization schema, the same name and address spelling in every source, independent third-party verification. If a panel forms, claim it through the link in the search result and edit some fields.

Should I open a Wikidata record for my business?

Only if you have an independent, verifiable source. Wikidata's notability rule expects the subject to be verifiable against a serious and publicly available source, or to match a valid identifier in a structured external database. A record opened for a business no independent source mentions ends up in a deletion discussion and is usually deleted; since the edit history is public, a fabricated reference is noticed quickly.

Which links belong in the sameAs field?

Living pages that genuinely belong to you or mention you: official social profiles, your professional or trade chamber member page, news and interview links that passed an editorial filter, listings in serious sector directories. Abandoned accounts and paid listing pages produce noise, not signal. Rather than raising the link count, make sure the brand name appears in exactly the same spelling on every listed page.

AI confuses me with another business — how do I fix it?

You cannot correct the answer directly; you can only change the sources feeding it and wait for a recrawl. Three things get done: sharpen the distinguishing information by filling the address, founder and founding date fields in the schema; write the district, founding year and founder's name into the about page in plain sentences; correct old or wrong directory listings one by one, marking closed branches permanently closed instead of deleting them.