You open ChatGPT between meetings and type your company’s name, half expecting a tidy paragraph about what you sell. Instead you get a confident description of a different company. Maybe it’s the Ohio plumbing supplier that shares your name, or a version of your startup from two pivots ago, or a polite shrug saying it doesn’t have specific information about that business. A prospect asked the same question last week. You will never know what they saw.
That moment is where most founders start asking the right question: how does ChatGPT know about my brand, and why does it sometimes get it wrong? The answer is less mysterious than it feels. ChatGPT pulls from two sources of knowledge, and before it can say anything about you it has to decide which “you” you are. That decision is called entity resolution, and it runs on signals you can inspect and fix. This guide covers both sources, the records the models trust, and a 10-point audit we call the NAMED score.
ChatGPT has two memories, and only one of them is current
The first memory is training data. Before a model ships, OpenAI trains it on a huge body of text: publicly available web pages, licensed material, and data from other sources. Everything the model knows without looking anything up comes from that corpus, frozen at a knowledge cutoff. Think of it as a library that stopped acquiring books on a fixed date. If your company launched, rebranded, or raised a round after that date, the library has no record of it, and nothing you publish today will change that edition.
Training data also has a density problem. A model doesn’t store a fact about you because it saw your homepage once. It forms a stable picture when the same facts appear across many independent pages: your name next to your category, your founder, your city, your product. A company with one website and three directory listings gets a faint, blurry impression, which is why the model may blend you with a better documented namesake or fill gaps with plausible guesses.
The second memory is retrieval. When ChatGPT searches the web, it fetches live pages and writes its answer from what it found, with citations. Retrieval is current, it can be influenced this week, and it is where most practical work happens. The two memories also interact. A model with a clear prior about who you are will interpret search results correctly, while a model with no prior is more likely to grab the wrong company’s page and stitch it into your answer.
OpenAI documents which crawlers feed which memory. According to OpenAI’s crawler documentation, GPTBot is the agent used “to make our generative AI foundation models more useful and safe,” which means training. Blocking GPTBot in robots.txt tells OpenAI not to use your pages for future models. That’s a legitimate choice, but brands that block it to protect content often don’t realize they are also opting out of the next library edition.
How does ChatGPT know about my brand when it searches live?
OpenAI launched ChatGPT search on October 31, 2024, first for Plus and Team subscribers and SearchGPT waitlist users. It reached logged-in free users on December 16, 2024, and on February 5, 2025 OpenAI opened it to everyone in regions where ChatGPT is available, with no account required. The launch post says the feature draws on third-party search providers plus content that news partners such as the Associated Press, Reuters, and Axel Springer supply directly.
OpenAI’s post does not name the search provider, but the history points one direction. At its Build conference in May 2023, Microsoft announced that Bing would become the default search experience inside ChatGPT, and independent testers, including Windows Latest in November 2024, found ChatGPT search results that matched pages from Bing’s first two pages of results. The practical takeaway is plain. If Bing can’t find, crawl, or understand your site, ChatGPT search is working with a handicap when someone asks about you. Verify your domain in Bing Webmaster Tools, submit a sitemap, and check that Bing’s index shows your current pages, not a stale snapshot.
OpenAI also runs its own search crawler. OAI-SearchBot, per OpenAI’s documentation, “is for search.” It surfaces websites in ChatGPT’s search features, and sites that block it won’t appear in ChatGPT search answers. A third agent, ChatGPT-User, visits pages when a person asks ChatGPT or a custom GPT to open them, and OpenAI says it is not used for crawling the web in an automatic fashion. OpenAI also notes that robots.txt changes take roughly 24 hours to register for search.
So when someone asks how does ChatGPT know about my brand in the live sense, the honest answer is this: it knows what OAI-SearchBot and its search providers can reach, ranked by whatever those systems consider relevant, filtered through the model’s prior understanding of who you are. Open your robots.txt today. It is common to find a company that blocked every bot with “AI” in its name and then wondered why ChatGPT search never cites it. Allowing OAI-SearchBot while blocking GPTBot is a valid configuration if you want search visibility without contributing to training.
Entity resolution is the step most brands never see
Before a search engine or language model can say anything useful about “Acme,” it has to resolve the string into a thing. Is this Acme the logistics software firm in Denver, the cartoon supplier from old Road Runner shorts, or the roofing contractor in Leeds? Search engineers call this entity resolution (or entity linking), and every modern system does some version of it.
Google made the idea public on May 16, 2012, when Amit Singhal, then its senior vice president of engineering, introduced the Knowledge Graph in a post titled “Introducing the Knowledge Graph: things, not strings.” At launch it held more than 500 million objects and more than 3.5 billion facts about those objects and the relationships between them. The phrase stuck because it described the shift in five words: search stopped matching keywords and started matching entities, each with a stable identity and a set of attributes.
Language models resolve entities in a looser, statistical way, but the inputs look familiar. The model weighs the context of the question (an industry, a city, a product name), the descriptions attached to your name across sources, and any structured identifiers it can find. When those inputs agree, the model locks onto one entity with confidence. When they conflict, because your LinkedIn says “fintech,” your Crunchbase says “payments,” your site says “revenue operations,” and a namesake has a Wikipedia article, the model hedges, merges, or picks the better documented candidate.
The fix is to make your entity easy to resolve. That means a small number of authoritative records that say who you are, markup on your own site that points to those records, descriptions that match everywhere, and independent coverage that repeats the same facts.
Which records do AI systems treat as ground truth?
Three sources carry outsized weight because so many other systems read from them. Wikidata, launched on October 29, 2012 and hosted by the Wikimedia Foundation, is a free, structured knowledge base where each entity gets a permanent identifier (a Q-number) and machine-readable statements: founding date, headquarters, official website, industry, founders. Search engines, voice assistants, and research datasets pull from it, and its content is openly licensed, which makes it a common ingredient in the data models learn from.
Wikipedia is the narrative counterpart. A Wikipedia article is one of the strongest signals of notability a brand can have, and that is the reason you can’t write one for yourself. Wikipedia’s conflict-of-interest guidelines discourage editing articles about your own company, and articles need significant coverage in reliable, independent sources to survive deletion review. Treat Wikipedia as an outcome of press coverage rather than a task. Wikidata has a lower bar, but the same rule about honesty applies: a well-sourced item with accurate statements helps, and a promotional one gets cleaned up or removed.
Google’s Knowledge Graph is the third record. OpenAI doesn’t say that ChatGPT reads it, so don’t assume a direct pipe. It matters for two other reasons. First, it shapes what Google’s own AI features say about you. Second, the Knowledge Graph and ChatGPT draw on overlapping public sources, so a messy knowledge panel is usually a symptom of the same inconsistent data that confuses every model. If you have a knowledge panel, claim it through Google’s verification flow and suggest corrections for wrong facts.
Check all three this week. Search Wikidata for your company name and note whether an item exists, whether it’s yours or a namesake’s, and whether its official website property points to your current domain. Then search Google for your brand and see whether the panel shows the right logo, founding year, and description.
Put your identity in markup with Organization schema and sameAs
Your own website is the one source you control end to end, so make it speak the language machines use for identity. Schema.org’s Organization type lets you state, in JSON-LD on your homepage or about page, your legal name, alternate names, logo, founding date, founders, address, and contact points. Most content management systems support it through a plugin.
The property that does the heavy lifting is sameAs. Schema.org defines it as the “URL of a reference Web page that unambiguously indicates the item’s identity,” and gives a Wikipedia page, a Wikidata entry, or an official website as examples. List your Wikidata item if you have one, your LinkedIn company page, your Crunchbase profile, your official social accounts, and any other profile that clearly belongs to you. Each link is a thread connecting your domain to a record a crawler may already trust. Together they tell a resolver that the company on this domain and the entity at a given Q-number are the same thing.
Keep the markup honest and tight. Only list profiles you control or that describe you, use the exact legal name your filings use, and put alternate names (an old brand, an abbreviation, a product name people confuse with the company) in alternateName so the model can map them to you. Then run the page through Google’s Rich Results Test or the Schema.org validator to confirm it parses. Markup with a syntax error is invisible.
Make every description say the same thing
Local SEO people have preached NAP consistency (name, address, phone) for years, and the principle carries straight over to AI. Every place your company appears is a vote on who you are. If the votes disagree, the model averages them, and averages are vague.
Start with the factual fields. Your company name should appear in one form, not “Acme,” “Acme Inc.,” “Acme Labs,” and “AcmeHQ” across four platforms. Your address, founding year, headcount range, and founders should match on your site, LinkedIn, Crunchbase, Google Business Profile, review sites such as G2 or Capterra if you sell software, Apple Business Connect if you have a physical location, and any industry directory.
Then fix the descriptive layer, which matters more for language models than for classic search. Write one canonical sentence that defines what you are, in plain category language: “Acme is a revenue operations platform for B2B SaaS companies, based in Denver.” Use that sentence, or a close variant, as the first line of your About page, your LinkedIn tagline, your Crunchbase summary, your press boilerplate, and your Wikidata description. Models learn associations from co-occurrence. If “Acme” and “revenue operations platform” appear together on twenty independent pages, that pairing becomes the model’s default answer. If each profile uses a different clever phrase, nothing sticks.
Watch for drift after a pivot. A rebrand is one of the most common causes of the wrong-company problem, because the old description lives on in hundreds of places you’ve forgotten about and in training data you can’t edit. Update everything you can reach, redirect old domains, and keep the former name in your schema’s alternateName so both strings resolve to the same entity.
Press coverage is the corroboration layer
Records and markup tell machines what you claim. Independent coverage tells them whether anyone else agrees. A brand that only describes itself looks, to a model, like an unverified claim. A brand that a trade publication, a regional business journal, and an industry podcast all describe the same way looks like a fact.
This is why press matters for AI visibility in a way it never quite did for links alone. Coverage in established outlets feeds both memories: those publications are well represented in the text models train on, and they rank and get cited in live search. OpenAI’s news partnerships, which include the Associated Press, Reuters, Condé Nast, Hearst, and the Financial Times, show which kind of source the company values enough to license.
The detail most PR plans miss is consistency. When you earn an article, the reporter will often lift your boilerplate or your founder’s description of the company. Make sure the boilerplate is your canonical sentence. Give reporters the same category language you use everywhere else, the same founding year, the same spelling of your name. Every piece of coverage then reinforces the entity rather than adding a new variant. At Instant Press we treat each placement as an entity signal first and a traffic source second, because a well-worded mention in a credible outlet keeps paying off in every model trained after it.
Quantity matters less than spread. Ten mentions on one site read as one source. Five mentions across five independent outlets in your category read as consensus.
Score yourself with the NAMED framework
To turn all of this into a working audit, we use a five-part model called NAMED. Each letter is one signal family, scored 0, 1, or 2, for a total out of 10.
N is Name consistency. Score 2 if your company name, address, founding year, and canonical description match across your site and at least five major profiles. Score 1 if the name matches but descriptions drift, and 0 if you find conflicting names or a stale brand still live.
A is Authority records. Score 2 if you have an accurate Wikidata item and either a Wikipedia article or a correct Google knowledge panel. Score 1 if one of those is in good shape, and 0 if none exist or the records describe a namesake.
M is Markup. Score 2 if your site publishes valid Organization schema with sameAs links to your Wikidata item and main profiles, plus alternateName for former names. Score 1 if schema exists but sameAs is thin or missing, and 0 if there’s no structured data at all.
E is Echoes, meaning independent coverage. Score 2 if at least three credible, independent publications have described you in the last two years using language close to your canonical sentence. Score 1 for one or two mentions or for coverage that describes you inconsistently, and 0 for none.
D is Discoverability. Score 2 if robots.txt allows OAI-SearchBot, your site is verified in Bing Webmaster Tools with current pages indexed, and your key pages render their text without depending on JavaScript. Score 1 if one of those is missing, and 0 if crawlers are blocked or Bing shows an outdated version of your site.
A score of 8 to 10 means models can resolve you with confidence, and the remaining work is staying current. A 5 to 7 usually means one weak family, often Echoes or Authority, is dragging down an otherwise clean profile. Below 5, expect ChatGPT to confuse you with someone else or decline to describe you, and fix Discoverability and Name first because they are the fastest wins.
ChatGPT knows your brand through two channels: a frozen library built from training data, and a live search layer fed by OAI-SearchBot and third-party providers. Both depend on resolving your name to one entity, and that resolution rests on authority records, clean markup, consistent descriptions, and independent coverage. Score the five NAMED signals, fix the lowest one first, and ask ChatGPT about your company again every quarter to see what changed.