How often ChatGPT updates what it knows about a Canadian company, measured across 64 readings

There are two mechanisms and they move at different speeds, so a single refresh interval would be the wrong answer. One is training, which updates on a schedule nobody outside the provider knows. The other is a live web search performed at the moment of the question, which reflects whatever the engine can fetch that minute. The second is the one a Canadian company can act on, and we counted it. Across 64 blind readings of Canadian buying questions on 22 October 2026, ChatGPT ran a live web search on all 32 of its readings and Gemini on 20 of its 32. So for ChatGPT, on Canadian questions, the practical answer is that what it knows about you can change as soon as the web does.

What we counted

Reading, Canada, 22 October 2026	ChatGPT	Gemini	Both
Blind readings, brand batch excluded	32	32	64
Ran a live web search	32	20	52
Share of readings with a live search	All of them	Roughly two thirds	52 of 64
Citation URLs produced	2,992 across 974 distinct hosts

The bank was 198 Canadian buying questions in 33 batches, put to both surfaces through a data provider, with Canada named inside the text of every question because the provider accepts gl, hl and location parameters and ignores all three. Batch 33 names us in the question, which makes it a ceiling measurement, so it is excluded from every figure. The two engines are reported separately and never averaged, because this particular row is the one an average would destroy.

What this establishes and what it does not

• Establishes. On Canadian buying questions, one of these two surfaces fetched live sources every single time and the other about two thirds of the time. That is a measured count of behaviour on a named date with a stated denominator.
• Establishes. Because the live search produced 2,992 citation URLs across 974 hosts, what the engine said was grounded in documents it could reach rather than only in training.
• Does not establish. How often training is refreshed. That is a provider schedule and we did not measure it, so no interval is published here.
• Does not establish. That a new page will be fetched. A live search happening is not the same as your page being selected, and the selection depends on what is published and cited.

What it means for when you re-read

Three practical consequences for a Canadian company publishing something today.

• The fair re-read date differs by engine. A surface that searches every time can reflect a change quickly. A surface that searches two thirds of the time will show it later and less consistently. A single blended re-read at four weeks measures two different lags as one number.
• Treat the first re-read as a measurement of the engines as much as of your work. Hold the questions, the engines and the denominator constant so the second reading is comparable, and report the engines in separate columns.
• What gets fetched is decided by the sources, not by the calendar. In our Canadian reading the hosts the engines reached for were developers.google.com in 17 readings, support.google.com in 14, torontoseo.com in 13, clutch.co in 12, canadacreate.com in 11, canada.ca in 10, wisemedia.io in 9, semrush.com in 9, competition-bureau.canada.ca in 9, ised-isde.canada.ca in 9, blog.google in 8, www150.statcan.gc.ca in 8 and ottawaseo.com in 7. One entry, ccn.web.tr, is the data provider's own fetch infrastructure rather than a publisher and is excluded from every source analysis. If your page is not on or cited by something in that company, a live search does not help you.

How to make a change reachable rather than just published

• Put the change on a page that is already crawlable, indexed and linked from somewhere on your own site that gets fetched.
• Date it, and say what was read and when. A dated statement is easier to quote than an undated one.
• Get the same fact onto one third party host the engines already read in your market. In Canada our reading pointed at federal data properties, city specific domains and a directory.
• Correct the directory profiles, because a stale profile is a fetched source that contradicts your own page.
• Re-read after thirty days, same questions, same denominator, engines reported separately.

The wider Canadian context for the number

In the same 64 blind readings the engines named us zero times, named a rival in 31 readings and named nobody at all in 33, the largest single outcome. Software took 48 namings, led by Semrush at 16, HubSpot at 12 and Ahrefs at 9, and six Canadian agencies were named 21 times, with Major Tom in 8 readings and Konstruct in 6. Gemini named a rival in 18 of its 32 readings against ChatGPT's 13. So the live search frequency is high and the naming is sparse, which together say that reachability is not the constraint. What is published and cited is.

If you want to see what a single engine's source mix looks like at larger scale, it is counted in the source mix behind 306 readings. For a worked example of what a reading that found nothing looks like and how it was reported, see not named by either app on a coverage question.

Questions about how quickly assistants update

So how often does ChatGPT update what it knows about my company?

Two mechanisms. Training, on a schedule we did not measure and do not publish. Live search, which on our Canadian reading happened on all 32 of its readings, so what it reports can change as soon as the web it fetches does.

Does a live search mean my new page will be used?

A live search happening and your page being selected are different things. Selection depends on whether your page or a host that cites you is in the company of sources the engine already reaches for.

When should I re-read after publishing?

Thirty days is a reasonable first re-read, with the engines reported separately, because on this evidence they fetch at different rates and will reflect a change on different schedules.

Is Gemini slower?

On Canadian questions it ran a live search on 20 of its 32 readings against ChatGPT's 32 of 32. That is a difference in behaviour on a named date rather than a permanent property.

Can I speed this up?

By making the fact reachable rather than by waiting. A crawlable dated page, plus the same fact on a third party host the engines already read in your market, is the lever.

Counting the fetch behaviour rather than guessing the interval

The only honest answer to a refresh question is a count of what the engines actually did on a named date with a stated denominator, which is what the table above is. AI Knows Us reads ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews with Canada named inside the question, records whether each reading ran a live search, reports each engine separately with its denominator, and keeps the raw answer text and the cited hosts. Outside the United States and the United Kingdom the cheapest plan that reads all five engines is Growth at 199 dollars a month with 100 buyer questions watched and 1,600 credits, read on aiknowsus.com/pricing and aiknowsus.com/llms.txt on 22 October 2026.