Good, fast, cheap is a reminder, not a law – and with AI the third corner quietly changes from “cheap” to “accurate.” Speed and cost collapsed toward zero. Checking did not. So the question is no longer which two you picked, it is whether the work included the checks needed to justify trusting it. Most of those checks you can run yourself, in about ten minutes per claim, with a timestamped archive of the page and two lines of notes next to the number. This piece shows you how, then tells you where that stops being enough.
- With AI the triangle is speed, cost and accuracy, and the failure mode is asking for fast and cheap, then assuming accuracy came included.
- Agreement isn’t verification. Confidence isn’t evidence.
- Two models agreeing is overlapping training data, not independent confirmation.
- A traced claim needs three things recorded: the page as it appeared, the date and time you looked, and whether the sentence is an observation or your reading.
- Do it yourself for a hypothesis. Buy the checking when the decision has a date and a reversal has a price.
The triangle changed shape, and nobody announced it
The old version you could feel. Good, fast, cheap, pick two – a printer, a contractor, an agency. The cheap job looked cheap. The fast job arrived unfinished.
The new version hides. Ask a model about your market and you get fast and cheap in the same breath, formatted well, footnoted, confident. The corner you gave up is accuracy, and accuracy does not announce itself as missing. A wrong answer looks exactly like a right one.
That is the whole trap: you asked for fast and cheap, then assumed accuracy came included.
One traced error, start to finish
Abstract comparisons do not move anyone, so here is the shape of the thing, and I’ll be straight about the limits of what I can show you on a public page.
The pattern we keep running into is this. A model produces a sentence with a number in it and a citation after it. The citation is real: the source exists, the link opens. But the number in the sentence is not the number on the page, or it is the number from an earlier version of the page, or it is a figure the source itself attributed to someone else. Nobody opened the page. The number went into a slide, the slide went into a board deck, and the first person to open the primary source was the one person you did not want opening it.
A citation proves a source exists. It does not prove the source says what the sentence claims, and it does not prove the page still said it last Tuesday. Three different questions, and only the third one has a date attached.
I want to give you the named company, the figure the model produced, the figure on the filing, and what the gap cost the decision, run live on a public claim so you can repeat every step yourself. I can’t publish that worked public example here yet, because the one specimen that needs no permission and no redaction is us: a full public diagnostic of Rathvane, built by the same pipeline, published with no PIN and a live Evidence Locker, is the piece we’re building first. Until that ships, I’m not going to stage a plausible-looking example and call it worked. What I can do is hand you the protocol we use, so you can run it on your own claim this afternoon, against a named model’s answer and the filing or pricing page it cited.
Two cases I can describe, both ours to tell. The first: a friend asked whether Rathvane could help with research for an industry report in the financial sector, so we put it to work, and it did the job: every claim stored in the Evidence Locker after verification, held there for future reference, nothing invented. Then I gave him the data and he turned his own agents loose on it. What came back had drifted. Rathvane found 19 errors in what his agents produced. The agents were fast and cheap. The checking is what nobody ran.
The second is our own near-miss, and it’s the reason the verification layer exists. On a demonstration build about a global brand, never sent to anyone, five of five competitor profiles carried wrong profitability figures, two with the sign flipped, every one of them citing a real filing. The citations opened. The numbers were wrong anyway. Our own checks caught all five before delivery. That is the exact failure this page is about, caught on our side of the line rather than in a board deck.
The do-it-yourself verification protocol
This is not a Rathvane product. It’s the manual version of the work, and it is worth doing on any claim you’re about to stake something on.
Step 1: open the primary source, not the summary
Follow the citation to the actual filing, listing, pricing page or press release. If the citation points at an article that points at another article, keep walking until you reach the thing nobody is paraphrasing. If the walk dead-ends at a 404 or a paywall, that claim is unverified, and it should be written down as unverified rather than quietly promoted.
Step 2: match the sentence to the page, word by word
Find the number on the page. Not a number near it. The number. Then check three things that break claims more often than outright fabrication: the unit (monthly or annual, gross or net), the date range the figure covers, and whether the source is reporting its own data or quoting someone else’s.
Step 3: archive the page with a timestamp
A link is not evidence, because the page can change under you. Save the page itself. A full-page screenshot works. A print-to-PDF of the page works, and it carries the URL and the date in the header if your browser is set to print them. A public web archive works, and has the advantage that a third party holds the copy. The file name should carry the date, because the date is the part that goes stale.
Step 4: hash the file, so you can prove it didn’t change
A content hash is a short string computed from a file. Change one pixel and the string changes completely. It is how you show, later, that the capture in your folder is the same capture you took. On a Mac or Linux machine, run shasum -a 256 yourfile.pdf in the terminal. On Windows, certutil -hashfile yourfile.pdf SHA256. Copy the output into your notes next to the file name. If anyone later questions whether the capture was altered, re-running the same command reproduces the string, or it doesn’t. No tooling to buy.
Step 5: write the four lines that survive into the board deck
Next to every claim in your own notes, record four things. Do this and the claim can be defended six weeks later by someone who wasn’t in the room.
- The claim, in the units the source used, not the units you wish it used.
- The source: what the document actually is, not the aggregator that mentioned it.
- The date and time you captured it, plus the file name of your archive and its hash.
- The register: observation, or your reading. If it’s your reading, write “our reading” in front of it, so the next person can discount it as they see fit.
That last line is the one people skip, and it’s the one that saves you. An observation with a capture behind it and an inference you drew are two different kinds of sentence. Mixing them is how a board stops believing the next slide.
Agreement isn’t verification
Two things get mistaken for checking, constantly.
The first is agreement. You ask one model, then another, they say roughly the same thing, and you relax. Models trained on overlapping public text agreeing with each other is not independent confirmation, it’s the same source in two voices. Our reading: consensus between tools tells you about the training data, not about whether your competitor raised prices in March. Independent confirmation means a second primary source, not a second model.
The second is confidence. Fluent prose with footnote markers reads like diligence. So does a junior analyst’s document with real citations and wrong numbers. Evidence suggests the expensive errors in a board deck aren’t the vague claims. They’re the precise ones nobody traced back.
How long does a capture stay true?
Here I have to mark the register carefully, because this page’s whole argument is that assertions need captures.
Our working rule is that a capture is good for about a week on a pricing page and about a month on a leadership page, and the card on each capture carries the date it stops counting as current. That rule is our judgment from doing the work, not a published study, and I’m not going to dress a rule of thumb up as a measurement. We have not published a re-check sample with counts, periods and change rates behind it. Until we do, treat the decay rule as our reading, and set your own expiry on anything volatile.
The practical version: pricing pages, pay-per-click landing pages and job boards move fastest. Leadership pages, filings and product documentation move slower. When a claim’s expiry falls before your decision date, recapture it. That is a five-minute job if you already have Step 3 in your habits.
When doing it yourself stops being the right call
The protocol above costs you nothing but time, and time is the thing you don’t have. Ten minutes a claim is fine for three claims. A competitive picture has sixty, and they interact.
So name the decision first, then pick the corner you can give up.
Sanity-checking a hypothesis before a Tuesday call? Ask a model, label it unverified in your notes, move on. Legitimate use, honestly priced at zero. Use it for exploration when the cost of being wrong is small and recoverable.
Setting a price, walking into a board meeting, responding to a competitor move, or raising inside 24 months? Accuracy is the corner you cannot give up, because a wrong number there isn’t an embarrassment, it’s a decision you can’t take back. That’s when a competitive intelligence diagnostic or a pricing strategy diagnostic earns its price. We’re slower than a chatbot and faster than a traditional intelligence firm or consulting agency, because the hours go into checking observations against primary sources rather than into drafting. AI is assistance in that loop, never the author.
FAQ
Does good, fast, cheap still apply to AI?
It applies with one substitution. Speed and cost both collapsed toward zero, so the corner under pressure is accuracy. The practical question is not which two you picked, it’s whether the work included the checks needed to justify trusting it. A quick answer can be correct and an expensive one can be wrong, which is why “pick two” is a reminder rather than a law.
How do I verify an AI-generated market claim myself?
Open the primary source rather than the summary, match the number word for word including its unit and date range, archive the page as a screenshot or print-to-PDF with the date in the file name, hash the file so you can prove it didn’t change, and write the source, the capture date and the hash next to the claim in your notes. Then mark whether the sentence is an observation or your own reading.
or call (888) 902-9777
- Every finding sits on a dated capture in the Evidence Locker.
- Every number traces to a primary source; estimates are labelled as estimates.
- A person who knows the client's name signs off before anything ships.
- Delivery in days, not months (3 to 15 working days depending on product and tier).
- Published pricing on every product, with a founding client rate through December 31, 2026.
- Satisfaction guarantee: if the work does not meet the published standard, the client does not pay.
- Ask Rathvane answers questions from beneath the surface of every deliverable, grounded only in the engagement's evidence.
The only number that matters
Cheap intelligence is cheap only if it’s right. Price the error, not the invoice: the quarter spent on the wrong lane, the pricing move you reversed, the board that stopped believing you.
Who wrote this
John Kirker founded Rathvane and built the method so every claim links to a dated capture of its source. If you’d rather see that on your own company than read about it, ask for one finding.
Reviewed by John Kirker, Founder - last updated November 16, 2026.
Sources for this article
- Operator brief: good/fast/cheap applied to AI; speed, cost, accuracy; 'asking for fast and cheap, then assuming accuracy came included'; 'pick two' is a reminder not a law; agreeme
- Operator brief: not cheap, not super fast, but faster than traditional intelligence businesses and consulting agencies; accuracy matters most.
- Corpus: a capture holds the page as it appeared, the prompt or query, the date and time, the source, and a content hash.
- Corpus: every statement is an observation with a capture behind it or a reading marked as ours; observations checked against primary sources; where a source is weak the deliverable
- Discovery transcript: capture good for a week on a pricing page, about a month on a leadership page; the card tells you the date it stops counting as current.
- Site requirements: more than eighty automated checks; people in the loop; AI-assisted never author.
- Pricing post: guarantee wording 'If we do not provide you value, you do not pay us a dime.'
- One Finding post: free, one checkable observation about your public surface.
Facts drawn from this site's own documented knowledge and first-hand experience.

