Field study

Field study: the ChatGPT API and the ChatGPT app recommend different businesses

We ran the same buying-intent questions through the ChatGPT app and the OpenAI API side by side, across four local verticals. The bare API — the way most tools query ChatGPT — recommended the same businesses as the app only about 8% of the time. Turning web search on roughly doubled the agreement and still missed two-thirds of the app's picks.

Updated July 27, 2026Reading time 6 min

What we did

There’s a widely repeated claim that AI-visibility tools querying the OpenAI API don’t measure what a customer sees in the ChatGPT app. We wanted our own numbers, on the questions that actually matter — local, buying-intent questions like “best [service] near me.”

So we ran a controlled comparison. We took 40 real buying-intent questions across four local service verticals — personal-injury law, family law, med spa, and cannabis retail, one market each in four U.S. states — and asked every question three ways:

We ran each question three times per surface, recorded the businesses each one recommended, and measured the overlap. (Four of the 40 were branded reputation questions about a specific business; those are set aside from the head-to-head, leaving 36 comparison questions.) Because the app and the browsing API arm run the same model, the differences below come from the product wrapper around it — a hidden system prompt, live browsing, memory — not from a different brain.

The headline

Across all four verticals, here’s how often each API surface recommended the same businesses as the app:

API surface Overlap with the app Recovered the app’s picks Named any local business
Bare API (no browsing) ~8% 17% 72%
API with web search ~16% 36% 97%
Search-native model ~9% 32% 100%

For reference, the app named a local business on 97% of these questions. “Overlap” here is the share of all the businesses recommended across both surfaces that both surfaces recommended — so 8% means that, of everything the app and the bare API named, only about one in twelve was named by both.

The single most telling number: the bare API declined to name any local business on more than a quarter of these questions — it fell back on generic advice or referral directories — while the app named specific businesses almost every time. When the model can’t browse, it often has nothing local to say, and says so.

What the bare model recommends instead

When the API can’t search, it recommends from memory. And memory favors the businesses with the loudest national footprint — the biggest advertisers and the national chains — over the local specialists the app surfaces after browsing. In two of our four verticals, the bare API’s recommendations overlapped the app’s by essentially zero: it named a completely different, nationally-familiar set, while the app named local, credential-verified businesses it found by searching in the moment.

Browsing narrows the gap. It doesn’t close it.

Give the same model a web-search tool and it moves toward the app: overlap roughly doubles, and it recovers about a third of the app’s picks. That’s the expected result — browsing is the single biggest thing the bare API is missing. But “a third of the picks” is still a different answer. The app’s wrapper does more than search; it also applies a long hidden system prompt and personalization that the API doesn’t, and those keep the two surfaces apart even when both are browsing.

A “search-native” API isn’t the shortcut either

We also tested a dedicated always-searching model — the kind of search-first API some tools lean on. It named a business every single time, but its overlap with the app was only about 9% — barely above the bare model. The reason is instructive: it browses widely and returns long, noisy lists, so it names more, but not the same. Always searching is not the same as matching what a careful consumer product recommends.

It even names businesses the app never surfaced

The gap runs both ways. On several questions, an API surface recommended a business that the app never named for the same question — including, in a few cases, the very business we were separately measuring, which the app had left out of its shortlist. That’s the point worth sitting with: the API isn’t a lower-resolution version of the app. It’s a different answer, and sometimes it’s confidently pointing at businesses the app quietly passed over.

The vertical spread

The gap wasn’t uniform. It was smallest in the vertical where the market’s most prominent name is both nationally famous and locally dominant — there, the bare model’s memory happened to line up with reality more often. It was near-zero in the verticals built on hyper-local specialists, where the bare model reverts to national chains it can recall. That’s a useful warning: how badly the API misrepresents your market depends on your market, and you can’t know which case you’re in without measuring.

What this means if a tool measures your “ChatGPT visibility”

Most automated tools query the API because it’s cheap and stable. Our data says that number describes a surface your customers don’t use — and on local, buying-intent questions, it diverges most exactly where it matters. If a dashboard tells you where you stand in “ChatGPT,” the honest first question is which ChatGPT — the browsing, personalized app a customer opens, or an API proxy that’s easier to script. That’s why we measure the consumer app, in a fresh session, and tell you which surface each result came from.

How to read this, honestly

This is a first cut, and we’ll say so plainly. It’s four verticals, one market and one focal business each, 36 comparison questions, three runs per surface — enough to establish the shape of the gap, not to put a decimal point on it. We measured the overlap of full recommendation sets, which is a stricter test than comparing only the top few names. And it’s our own study, run to inform our own method — read it as directional evidence that the API and the app are different products, which is the same conclusion the broader, cross-engine case reaches from the vendors’ own documentation.

If you want to see it for your own business — measured on the app a customer actually uses, across all six engines — that’s the free report.

Common questions

Does the ChatGPT API return the same recommendations as the ChatGPT app?
In our testing, no — not close. Across four local verticals, the bare OpenAI API (no web search) recommended the same businesses as the consumer app only about 8% of the time, measured as the overlap between each surface's recommended set. Enabling web search roughly doubled the agreement, to about 16%, but it still missed roughly two-thirds of the businesses the app recommended. The API and the app share a model family, but they behave like different products.
If an AI-visibility tool measures ChatGPT through the API, is that a problem?
It depends what you're deciding. The API is cheap and scriptable, so many tools use it — but our data shows its recommendations diverge sharply from the consumer app your customers actually use, especially on local questions. If a tool reports your 'ChatGPT visibility' from the API, treat it as directional, and ask which surface it read. We measure the consumer app on purpose.

The report is free

See your visibility on the surface customers actually use.

Six engines, the consumer interface, every run on file — in 48 hours.

Get your AI visibility report