How assistants pick the GPU prices they quote
Nobody outside the labs knows how assistants weigh sources. A price is easier to quote when it is readable as text, tied to its page and dated.
What anyone can honestly say about it
Nobody outside the labs that build assistants knows how they weigh one source against another, and anyone selling certainty about it is guessing. What can be reasoned is narrower. An assistant can only quote a GPU price it can read as text, connect to the page it came from, and place in time. A rate that fails one of those tests is harder to quote than a rate that passes all of them.
That is a design decision on our side, not a finding. We have not measured which pages assistants cite, we have no view into their ranking, and we will not sell a tactic as a guarantee. What follows is the reasoning, how we applied it to our own price tables, and what a GPU provider can take from it for its own page.
Extractable: the price survives being lifted out
A rate in an HTML table cell, next to the full name of the card and the word hour, can be read by any program that reads text. A rate drawn on a canvas, set inside a screenshot of a table, or shown only after a buyer picks a region from a dropdown may never reach that program at all. Content behind interaction is one way a pricing page can hide its prices from machines without meaning to.
The sentence around the number matters as much as the cell. A banner that says "from" and a figure tells a reader nothing about which card, which commitment or which unit. A line that names the H100 SXM 80GB, on demand, per GPU per hour, can be quoted on its own and still be true. We apply that test to our own answer sentences: lift one out, read it cold, and see whether it misleads.
Attributable: the price arrives with its page
A figure that carries its origin is easier to repeat with credit. Every rate on the GPU price table sits beside the URL it was read from, so a number taken from our table brings its source along. We hold 464 current rate observations from 48 providers, and none of them is shown without that link.
For a provider, attribution starts with being the original. If the cleanest statement of your H100 rate is on a reseller page or in a forum thread, that is the version lying around to be quoted. Your own page should be the clearest copy of your own price, using the card name exactly as buyers type it into a search box.
A directory listing is a second copy, never a substitute. We read your page and link back to it, which also means a wrong figure on your page becomes a wrong figure in two places until somebody fixes the first one.
Fresh: the price has a date on it
GPU rates move, and a price with no date leaves every reader to guess its age. We hide a rate older than 21 days rather than show it stale, and the page on how our figures are gathered explains the rule. A gap in our table means no current rate is on record, which is a different statement from saying the card is not sold.
A provider can put a line reading "prices last changed" above its table for nothing. It is the cheapest freshness signal there is, provided it is honest. A date that rolls forward every morning whether or not a price moved is worse than no date, because the first buyer who notices stops trusting the table under it.
Freshness also decides what a buyer sees on a page like ours between reads. If you cut your H100 rate on a Monday and we last read your page the week before, the older and higher figure is what sits beside your name until the next read. A dated page makes the difference obvious to anyone who clicks through, and it gives a buyer a reason to believe the newer number rather than splitting the difference.
A ten-minute test for your own pricing page
Open your pricing page with JavaScript switched off, or view the raw source, and search for your H100 rate as digits. If the number is not in the HTML that the server sends, a crawler that does not run scripts never sees it. The check takes a minute and needs no tools beyond a browser.
Then copy the line that holds the rate into an empty document and read it without the page around it. Does it name the card and memory size, the unit, the commitment and the currency? Is there a date anywhere near it? If a colleague reading only that line would misquote you, a program reading only that line will too.
Last, search for your own card and price in a few assistants and see what comes back and where it says the figure came from. Treat the answer as an anecdote, not a measurement, since the same question asked twice can return different sources. It is still a quick way to find an old rate that has outlived the page it was on.
Where this reasoning stops
It says nothing about which assistant reads which page, how often, or with what weight, and nothing about whether a page that passes every test is quoted ahead of a competitor. Assistants change often and their makers publish little about how sources are chosen. Treat any claim to the contrary with suspicion, including this post if it reads that way to you.
The same three properties help a buyer who never opens an assistant. Someone comparing the H100 SXM 80GB across 13 sellers wants the card name, the unit, the source and the date. Fixing your page for that person is the part you control.
What we could not answer
- ⚑We have no data on which pages any assistant cites or how often, so this is reasoning about what makes a price quotable, not a measurement of what gets quoted.
- ⚑Assistants change their behaviour without notice, and nothing here can say how long any of it holds.
What this rests on
infratoapp, “How assistants pick the GPU prices they quote”, read 27 October 2026. current GPU rate observations, each with its source page: 464. https://www.infratoapp.com/blog/how-assistants-pick-the-gpu-prices-they-quote
Questions people ask about this
How does ChatGPT decide which GPU prices to quote?
Its makers have not published how it weighs sources, so nobody outside can say. What can be reasoned is that a price must be readable as text, tied to a page and dated before it can be quoted with any confidence. Pages that hide rates behind scripts, images or clicks make that harder.
Can a GPU provider guarantee an assistant will cite its prices?
No. No tactic guarantees a citation or a position, and nobody outside the labs knows the weighting. A provider can make its own page the clearest, most current statement of its own rates, which helps human buyers in any case and removes the obvious reasons a program would skip it.
Why does a price need a date to be useful?
GPU rates change and an undated figure could be from any month. A reader, or a program summarising for one, cannot tell whether it still holds. A visible date that changes only when the prices change lets the reader judge, and an honest date is cheap to add above any table.
These render as FAQPage structured data · answers are filled from the same live figures
Read next
A price with no date could be from any month. Show when your rates last changed, in text and in markup, and change that date only when a price moves.
A benchmark result is checkable only with its method beside it: the hardware, software versions, model, precision, batch size, sequence lengths and runs.
llms.txt is a plain-text index of your key pages. Google has said its search does not use it, some tools may read it, and it takes about an hour to write.