Published September 29, 2026 · Last updated September 29, 2026 · Completeseo's guide was updated September 28, 2026 — this page is re-checked when the test numbers change.

When ChatGPT names my business, is it remembering me or reading the web right now?

Direct answer: It can be either — and the answer you get decides your next move. ChatGPT talks about your business from two layers: the model's trained memory (what it absorbed about you from the public web before training ended) and live web search (what it retrieves right now). You can find out which layer is carrying your name in about two minutes. Open a new chat, turn web search off, and ask: "Without searching the web, what do you know about [your business name] in [your city]?" If it still knows you, that's the memory layer talking. If it only knows you with search on, you're riding on live retrieval. The fix differs because the clocks differ: search-based answers react to what you publish once it's indexed — that side moves first, in weeks — while memory-based answers reflect your reputation and change slowly, over months. (Test protocol: completeseo.com, "How AI search works for local businesses," updated September 28, 2026; the two-clocks framing was first put in buyer voice by an azbigmedia.com FAQ on September 17, 2026.)

The two-minute test

The protocol comes from completeseo.com's "How AI search works for local businesses" (updated September 28, 2026):

Memory answer. Open a new chat and ask, "Without searching the web, what do you know about [your business name] in [your city]?" — "The first answer is roughly what the model holds without searching."
Live answer. Then ask it to search the web for the same business — "the second is what it finds live."

Diff the two answers. Where they disagree, that's your gap — and where they agree, that's your solid ground.

The completeseo team ran this at scale, and the number should reset your assumptions: with its web search setting off, 95.5% of 3,303 answers still carried ChatGPT's card of businesses. The card — the named-business list your customer actually reads — is mostly drawn from memory, not from live retrieval. So "remembering you" is the default, not the exception. If a customer asks "who's the best water damage contractor in Tacoma" and your name appears, most of the time that's the model's long-term memory speaking, not a fresh look at your website.

What the two layers actually are

Layer one: training memory — the model's long-term knowledge of you. Before a model is released, it trains on an enormous slice of the public web: your website, your reviews, directory listings, "best of" roundups, local news, forum threads. What survives training is a compressed impression of what the internet said about you. Nobody publishes a page "for the training set" — you influence this layer only by being talked about widely and consistently across the web. And it moves slowly. The model's knowledge of you refreshes on the training-release cycle, not on your publishing schedule. Think months, not days.

Layer two: live web search — what ChatGPT retrieves right now. When the browsing/search tool is switched on, ChatGPT goes and fetches current pages — your site, review platforms, directories — and grounds its answer in them. This is the layer that reacts to what you publish once it's indexed: new service pages, fresh reviews, corrected listings. It moves in weeks.

Most recommendation answers are a blend. The candidates — the shortlist of names on the card — lean heavily on memory (remember the 95.5%). Live search then confirms details (hours, ratings, availability) and fills in gaps for names the model doesn't know. That division of labor is why you can fix your website on Monday and still hear the same old answer in August.

Why the distinction decides your next move

Say you're a restoration contractor in Tacoma and you run the two-minute test. There are four outcomes, and each one points to a different fix:

Memory knows you, search confirms you. You're solid on both clocks. The play is maintenance: keep publishing, keep reviews coming, keep listings consistent so the next training cycle has even more evidence of you. This is the position you defend, not the one you fix.
Memory knows you, search shows nothing new. Your reputation carries you — past reviews, old mentions, a well-established name — but fresh content isn't being picked up. The search side of your presence is stale or blocked. Check whether AI crawlers can actually reach your site, whether your new pages are indexed, and whether your listings are current. You have weeks-to-fix work here, not months.
Memory doesn't know you, search finds you fine. Typical for a newer shop, a recent rebrand, or a company that only recently started getting reviewed. The live layer works — a customer asking today sees you. But the memory layer is empty, which means the names that survive into the training snapshot are still your competitors'. This is normal, not broken. The fix is to keep feeding the search side (reviews that name the job, mentions on directories and local roundups, answer pages on your site) so the next training cycle has material about you. Patience is part of the strategy on this one.
Memory doesn't know you, search doesn't find you. The whole pipeline is broken. Start with the live layer first: can AI crawlers even get into your site? Are your pages indexed? Is your business findable on the directories and review platforms the search tool retrieves? Only after the search side answers correctly does it make sense to think about the memory side. (See what ChatGPT actually understands about your business for the access-first pre-audit.)

Notice what didn't appear in any of those four: "panic and rebrand your services page." The two-minute test exists precisely so you don't fix the wrong layer.

What to actually do this week

Run the two-minute test. New chat, search off, "Without searching the web, what do you know about [your business name] in [your city]?" Then the same with search on. Write down which layer knows you. This is your baseline — everything below hangs off it.
Fix the search side first — it's the fast clock. Publish one page that answers a real buyer question in your trade and city in quotable sentences (what a job costs locally, how long it takes, what happens on the first visit). Confirm AI crawlers aren't blocked from your site. Check that your listings are accurate and current. Search-based answers react once content is indexed, so this side moves first.
Start feeding the memory side for the long clock. Reviews that name the job ("they dried out our flooded crawl space in two days") instead of just "great service." Mentions on directories and local roundups. A consistent name, address, and service area everywhere (see how your Google Business Profile feeds AI answers). Memory reflects your reputation — and reputations change slowly because the underlying evidence changes slowly.
Re-run the test monthly. The search side should move within weeks of you publishing. The memory side will take months. If the search side doesn't move at all, the access layer is broken — crawlers blocked, pages not indexed — and that's where your effort goes. Fold this into your monthly AI-search visibility check.
Write down which layer your competitors live in. If a competitor is named with search on but gone with search off, they're renting visibility from the live web — beatable with better content. If they're in the memory layer, they've been building reputation for years and you outrun them by becoming the best-documented answer in your trade, not by outposting them for a month. (For the stable-rival case, see why ChatGPT recommends your competitor instead of you; for the flip-flopping case, see why a recommendation vanishes week to week.)

The line to remember

One answer, two clocks. When ChatGPT names your business, you're looking at either what it remembers about you or what it's reading right now — and the two-minute test tells you which. Fix the reading side this month; the remembering side takes a season. Do them in the wrong order and you'll wait months for a change that could have taken weeks.

FAQ

Q: Is this the same as ChatGPT's "memory" feature — the one that remembers things about me as a user?
A: No — different memory entirely. ChatGPT has saved memories about the user (preferences, details from past chats) that can steer an answer. This page is about the model's trained memory of your business — what it knows about you as a company, independent of who's asking. The two-minute test isolates it by using a new chat with web search off. They're separate layers, with separate fixes, and mashing them together is how the confusion starts.

Q: The test knew my business with search off — does that mean I'm safe?
A: It means you're known, which is necessary but not sufficient. The 95.5% figure from completeseo shows almost every local answer carries a memory-drawn card — so being in memory puts you in the pool, but the pool is big. Whether you're named in a given recommendation still depends on reviews, ratings, mentions, and how the model weighs your reputation against competitors'. Known beats invisible. Known and recommended are two different things.

Q: I published new pages two weeks ago and the answer hasn't changed. What's wrong?
A: Probably nothing — you may be watching the wrong clock. Search-based answers react once content is indexed, which takes weeks; memory-based answers reflect your reputation and change slowly, over months. If you published two weeks ago, re-run the test with search on: that's the layer that should move first. If even the search-on answer hasn't moved, the problem is access — crawlers blocked or pages not indexed — not the calendar.

Q: Which fix comes first — the memory side or the search side?
A: The search side, always. It's faster (weeks vs months), it's fully in your control (your site, your listings), and it builds the evidence the memory layer eventually absorbs. A contractor who fixes the search side this quarter is building the reputation the model remembers next year.

Q: Does the customer see the same answer either way?
A: Not necessarily. A quick one-shot question leans on memory — that's the 95.5% case. A longer conversation with browsing on gets live-grounded answers where current reviews, hours, and availability get confirmed. Same customer, same question, different session settings — different answers. This is why you want both layers healthy: you don't get to choose which one the customer meets.