ChatGPT searches for the names it already knows. Make sure it knows yours.

Cloud AI cites who it can see. New research adds a harder part: it also searches for who it already remembers.

Yesterday we looked at how ChatGPT searches inside the official source's own site. This one is about the step before that. When someone asks ChatGPT to compare options, the searches it runs in the background often already have brand names in them. Names the person never typed. Names the model brought with it.

The data comes from Graphite, a growth agency that sells AEO services. On October 16, 2026 it published Parametric Knowledge Optimization, and Search Engine Journal covered it on October 26. "Parametric knowledge" is just what the model learned in training, before it searches anything. Graphite calls it "the model's fuzzy recollection of the internet."

The short version

• ChatGPT's searches aren't blank. Graphite looked at about 957,000 of ChatGPT's background searches ("fan-out queries") behind 22,953 comparison prompts that named no brand. 59.6% of those searches contained a brand anyway.
• Many of them were also site: searches. 22.7% used the site: operator, which limits a search to one website. That lines up with what we covered yesterday.
• Being on the cited pages wasn't enough. Brands mentioned on at least 15% of the pages ChatGPT cited still reached only 11% median visibility with search, when the model rarely named them without search (0–25% of the time).
• What the model remembers and what it finds both mattered. Visibility rose with each. Graphite says the same pattern showed up in two replications.
• Memory updates slowly. For models released since July 2026, Graphite counts four to five months between OpenAI's stated knowledge cutoffs and release dates. Anything you publish today won't be in a model that's already out.
• For a business, the job has two halves: be on the pages AI search finds now, and be described the same way across enough of the web that the next model knows your name.

Why a brand name in the search matters

When ChatGPT decides to search, it doesn't type your customer's question into a search box. It writes its own searches. Graphite's point is that those searches are often seeded with "candidate answers," the brands the model already expects to be good options. In Graphite's words: "Including a brand in a fan-out query increases the chance that pages featuring that brand will be retrieved."

That turns into a loop. The model searches for names it knows. Pages about those names come back. The answer cites those pages. If your name was never in the search, you're relying on showing up in someone else's page about someone else.

And even then, the study suggests it's an uphill fight. A model can read a page that mentions you and still not name you, if it didn't know you to begin with. Graphite also points to earlier academic work showing models sometimes lean on what they learned in training even when the page in front of them says something different.

Search is used more, but not always

In Graphite's fixed set of comparison prompts, the share of logged-out ChatGPT answers with a citation rose from 56.6% in the week of August 3, 2026 to 88.5% in the week of September 28. Graphite says plainly that these rates have changed over time and could go down. And when ChatGPT doesn't search at all, the answer comes entirely from memory. There is no page to be cited from.

What this means for a business

You can't edit a model's training data. You can control the conditions it learns from and the pages it finds. Five of them are in your hands:

1. Let the search crawler in. OpenAI's crawler documentation says sites opted out of OAI-SearchBot "will not be shown in ChatGPT search answers, though can still appear as navigational links."
2. Make the training decision on purpose. OpenAI's same page says GPTBot crawls content "that may be used in training," and that each setting is independent: you can allow search and block training. Graphite's advice is to allow AI labs' training crawlers if you want future models to know you. That's a real trade-off, not a default. Our robots.txt Field Guide walks through which token does what.
3. Check the layer under robots.txt. Graphite notes that allowing crawlers in robots.txt "alone is not sufficient" if a CDN or firewall rule blocks them, and suggests serving key content in server-rendered HTML, since training crawlers may not run JavaScript.
4. Describe yourself the same way everywhere. Same business name, same services in the words customers use, same town, same kind of customer. Graphite's hypothesis is that inconsistent descriptions make a model less confident about naming you. Put dates on anything that changes, like hours, offers, or service areas.
5. Earn mentions on many different pages. A search answer usually cites a handful of pages (Graphite says 4–7). Training data is vast and gets de-duplicated. One great page can help in search. Being known takes mentions across many distinct sites, not the same blurb pasted fifty times.

A brand-memory check (about 20 minutes)

1. Write three questions a customer would ask when comparing options in your category and town. Don't use your name.
2. Ask ChatGPT each one in a fresh chat. Note who gets named, and whether it searched (look for sources).
3. Then ask it directly what it knows about your business: what you do, who you serve, where. Graphite suggests this to tell "doesn't know you" apart from "knows you but rarely mentions you."
4. Compare its description to your own site, your Google Business Profile and your top directory listings. Fix the places that disagree.
5. Confirm OAI-SearchBot isn't blocked, and decide on GPTBot deliberately.
6. Re-check when a new model ships, and expect months, not weeks. Monitoring tells you what changed. It doesn't fix the description.

What this is not (honesty block)

• Not OpenAI's explanation. OpenAI hasn't published how ChatGPT writes its background searches or why brand names appear in them. These are observed searches, not documented rules.
• Not cause and effect. The study shows search visibility rising alongside "no-search" visibility. SEJ notes the write-up doesn't describe controls for how well known each brand already was. Famous brands may simply win both.
• Not a local-business study. The prompts were entity comparisons ("What is a good free scheduler?" is Graphite's example), collected logged out. Nobody here tested "best roofer in Provo." The mechanism is shown; rates for local services are unknown.
• A vendor study. Graphite sells AEO services and is proposing a new practice (PKO). Brand labels in the 957,000 searches came from an AI model (Claude Haiku 4.5), not human review. The method is exploratory and not peer-reviewed. SEJ also notes the write-up doesn't fully explain how no-search answers were separated in the main comparison.
• Not a training guarantee. Labs don't publish their training sets. Allowing GPTBot doesn't mean your pages get used, or when.
• Not a timeline promise. The four-to-five-month gap is Graphite's count from vendor-reported cutoffs. It's the minimum delay, not a schedule.
• Citation rates move. The 88.5% figure is one tool's fixed comparison prompts in one week. Graphite says it could go down.

Related New Reward Field Guides

• ChatGPT searches inside the official site. Make sure the answer is on yours.
• Cloudflare is helping ChatGPT find more pages. Found is not cited.
• robots.txt is a door, not a usage policy
• Reddit didn't vanish. ChatGPT stopped treating threads like source material.

Want the fast read first?

The free AI-visibility score is a quick check: email only, no credit card, a 0–100 read across the engines we track. It won't teach a model your name. It will show whether AI can see your business clearly.