ChatGPT says it has no real-time data in 1,513 answers, Gemini in 10

Short answer: ask ChatGPT for a plumber and almost one answer in two warns you, in its own words, that it cannot see anything current. It says it has no real-time data, or that its knowledge stops at a past date, in 1,513 of its 3,059 answers across eight countries. Gemini says it in 10, Perplexity in 7. And in 1,527 of those 1,530 answers, the assistant names a business anyway.

You ask for the best electrician in your city. ChatGPT gives you five names, then adds a line you might skim past: as of its last update, it cannot provide real-time recommendations. In 296 answers it puts a date on that update - October 2023 in 281 of them, October 2021 in 15. So the business it just recommended is the one it remembers, one to three years earlier: the address it held then, the hours it held then, the owner it held then. Gemini and Perplexity, asked the same questions in the same cities, almost never say this. ChatGPT says it in half its answers, and recommends you regardless.

It recommends you to the customer, then says it is working from a version of you that may no longer exist.

What was measured

We went back to the eight Waseit studies: two buyer questions, the best providers of a trade in a city, then the most recommended and why, across 89 cities in France, the United Kingdom, Brazil, the United States, Canada, Australia, New Zealand and Singapore, between 30 July and 16 October 2026, each in the country's language. That is 3,059 ChatGPT answers (gpt-4o-mini), 3,058 from Gemini, both answering without web search, and 1,364 from Perplexity, which does search the web and was only asked in three countries. We counted a confession only when the model said, in the first person, that it has no real-time or live data, or that its own knowledge stops at a past point - "I can't provide real-time recommendations", "as of my last update". We did not count a platform's offer of real-time data to the customer, nor advice to the customer to check current hours; those are a different thing. Our first detector was English-centric and missed the Portuguese and French forms; two reading passes corrected it, and the count is a floor.

What exactly does ChatGPT say, and does it still name a business?

Two forms, often in the same answer. In 1,448 answers it says it has no real-time data: "I can't provide real-time recommendations". In 292 it dates its own knowledge: "as of my last update", "based on prior knowledge". Either way it keeps going. In Los Angeles, asked for an electrician, it writes: "While I can't provide real-time recommendations, I can suggest some well-regarded electricians in Los Angeles based on prior knowledge" - and lists them. The one difference is what the model does next. ChatGPT names a shortlist in all but a handful of its 1,513 answers. Gemini and Perplexity, in the 17 answers where they say the same, do the opposite: they decline to rank and send the customer to a platform instead. In Fortaleza, Gemini writes that because it has no real-time database of customer reviews, the best way is to use trusted service platforms. For them the confession is a reason not to choose; for ChatGPT it is a footnote under a choice already made.

What it means if you are the business

The assistant that names businesses most often is also the one that most openly admits it may be describing a you that has changed. That matters in a precise way. If ChatGPT recommends you from an October 2023 snapshot, it is recommending the address, the hours, the phone number and the owner it held then - and if any of those moved, the customer it sent arrives at the wrong door. Two practical consequences. First, the facts a machine gets wrong with age - your name, address, hours, phone - have to say the same thing everywhere a machine can read them, so that an old memory and a live search still point to the same place. Second, nothing you write makes ChatGPT's training fresher; what you can do is be consistent and current everywhere it, or a customer, will look next. These are reasonable steps, not levers with a measured effect, and we did not test whether they get a business named more often. The open question is the one you cannot see from here: when ChatGPT recommends you from memory, is the version it remembers still you? Yesterday we looked at where the assistant sends your customer for recourse: AI says where to check a business 218 times, where to complain once.

And your business, what do the assistants say?

The free Waseit audit asks ChatGPT and Gemini the questions your customers actually type, in their language and for your city, and reads the answers one by one. It asks ChatGPT the same way this study did - through its programming interface, where it answers from memory - so what you see is exactly the remembered version this article is about. In about 60 seconds you get a score out of 100, the competitors named instead of you, and the questions where you do not appear. Free, no account and no card; an email address only unlocks the competitor names and the action plan. It is a dated snapshot, not a ranking: the answers change from one day to the next, and we promise no position. But it tells you what the machine says about you today, which is the one thing its own disclaimer admits it cannot vouch for.

What this does not prove

• These are model answers through their programming interfaces, with no user location; ChatGPT and Gemini answered without web search, Perplexity searched and was only asked in three countries. ChatGPT here is gpt-4o-mini, not necessarily the model behind the app, which can browse the web and would not give this disclaimer; the models may have changed since. Each question was asked once, between 30 July and 16 October 2026.
• The count is a positive count of explicit first-person phrases. A confession worded differently could have been missed, and two reading passes only lowered that risk - 1,513 is a floor, not a ceiling. The two forms overlap: 227 answers both deny real-time data and date their knowledge, so they are not added together.
• The disclaimer is almost entirely an English and Portuguese behaviour. In French we found it in 61 answers; the near-absence elsewhere for Gemini and Perplexity is a count of words, and a differently phrased admission our detector did not catch could exist. We read the Gemini and Perplexity near-misses one by one: the ones we set aside were a platform offering real-time data to the customer, not the model speaking about itself.
• Saying "as of my last update" is not proof that the recommendation is wrong. Many of the businesses named are long-established and have not moved. What the disclaimer establishes is that the model cannot itself tell which ones have - not that any particular one has.
• We did not measure whether being consistent and current gets a business named more often, nor whether the app, searching the web, corrects what the API remembers. We measured what the models say, not what is true about each business.