TL;DR
Open each link, then search the page for the exact number or quote the AI attributed to it. A link that loads proves nothing, because invented links often point to real sites. Then treat what fails differently depending on what it is, and remember that a calculation is only as reliable as its weakest input.
Open every link the AI gives you, then look on that page for the exact number or sentence it was cited for. If it is not there, the source does not exist, even when the link loads perfectly. What you do next depends on what failed: a number can survive as an estimate, a quote cannot.
Why AI sources need checking at all
Language models produce text that looks like research: a figure, an organisation, a year, a link. The format is convincing. The content is not guaranteed.
This is not a rare glitch. On Vectara's hallucination leaderboard, which measures how often models add facts that are not in a document they were asked to summarise, rates in 2026 ranged from 1.8% to 24.2% depending on the model. That is the easy case: the document is right there in front of the model. When it has to go and find sources on its own, there are more ways to get it wrong.
If you are a solo founder using AI to research your market, your customers or your competitors, every number you keep ends up in a decision: a price, a target, a plan. A made-up figure in your research becomes a made-up figure in your plan.
Why "the link works" is not a check
The intuitive test is to click the link. If it opens, the source is real. That test lets almost everything through.
An invented link usually does not look invented. The most frequent pattern is the right domain with a made-up path: the AI knows a national statistics office or a well-known research firm publishes on the topic, and writes a plausible URL on their site. Sometimes it 404s. Often it redirects to the homepage, or the AI links straight to a homepage or a section page. Either way, you land on a real, official-looking site that says nothing about your number.
So the check has two steps, and the second one matters more:
- Does the link open? A dead link is an easy fail.
- Does the page say what the AI said it says? Search the page (Ctrl+F or Cmd+F) for the number, and for the key words around it. If it is a PDF report, search the PDF. If the figure is not there, the link is decoration.
When step 2 fails, look for the original yourself. Search the organisation's name with the figure. If you find the real page, keep it and replace the AI's link with the one you actually read.
Three kinds of material, three different fates
Not everything an AI cites deserves the same treatment when the source does not hold up. Sort what you got into three piles.
Numbers. A market size, a price, a share of people who do something. If you cannot find the source, do not delete the number by reflex. It may still be roughly right, and you may have nothing better. Keep it, but remove the link and relabel it: "estimate, source not found". It stops being a fact and becomes an assumption you know you are making.
Quotes. "As one freelancer put it, I lose two hours a week on invoices." If you cannot find that sentence on the cited page, delete it. A quote presented without its page is a sentence someone wrote while claiming a real person said it. There is no such thing as a "probable quote".
Places. A forum, a professional group, a community where your customers gather. If the link is wrong, the name is often still real and useful. Keep the name, drop the address, and search for it yourself.
The logic is simple: keep what is still useful once you are honest about it, delete what only works if it is true.
A worked example: twelve sources, one calculation
Say you asked an AI to research the market for a booking tool sold to independent bakeries. It came back with twelve sourced claims. All the figures below are invented for the example.
You check each one:
- 3 links are dead. Easy fails.
- 4 links open, but the page does not contain the figure. Two are homepages, two are articles on a related topic.
- 5 links open and the page says what the AI said.
So 5 out of 12 survive: 5 ÷ 12 ≈ 42%. More than half of the "research" was unsourced, and four of those seven failures would have passed the "does it open" test.
Now look at the calculation you were about to build on it:
8,000 bakeries × 40% ready to pay × €360 a year = €1,152,000 a year
The count of 8,000 bakeries comes from an official register and you found it on the page: solid. The €360 a year is what they already spend on a comparable tool, and you found it too. But the 40% came from one of the four links that opened on a page without the figure.
That one input sets the reliability of the whole result. The total is not "mostly sourced". It is an assumption, because one of its three factors is. And it moves a lot:
8,000 × 20% × €360 = €576,000 a year
8,000 × 10% × €360 = €288,000 a year
Same register, same price, and the answer moves by a factor of four depending on a number nobody can trace. The honest way to write this is: "€1.15 million a year if 40% are ready to pay, which is an assumption." Then go and test the 40%, because it is the input that carries the most weight.
This is the rule worth keeping: a combined result is only as reliable as its weakest input. Multiplying a solid number by an assumption does not make the assumption more solid. It makes the solid number look like it supports a conclusion it does not.
What this method does not cover
Checking that a page exists and contains the figure tells you the AI did not invent the citation. It does not tell you the figure is good. A real source can still be old, measure something slightly different from what you need, or cover the wrong country. If the number feeds a market estimate, also check what it counts and where. Sizing the zone you actually serve is part of that.
It also does not fix the questions you asked. If you asked the AI to "find the market size" for a segment that is too vague, every source can be real and still be useless. Defining who actually pays for your app comes before any of this.
And it does not scale well by hand. Twelve sources is an afternoon. A full market study with fifty claims is a weekend, which is why most people stop at clicking the links.
How Banast does it
In a Banast report, every number shows its source and a reliability label. The source that counts is not the link the model wrote: it is the page the search actually consulted. The model's link is matched against those pages, and when it matches only by domain, Banast keeps the address of the page it really read, never the one the model wrote. A second check then confirms the link still responds.
When a figure cannot be matched, it stays on screen but loses its link and becomes "Stated", with low reliability. A customer quote that cannot be matched is removed. A place keeps its name without its address. And a result computed from several figures takes the reliability of the weakest one.
FAQ
The link works. Isn't that enough?
No. A common failure is a real domain with an invented path, or a link to a homepage or a section page that never mentions the figure. The test is whether the page you open contains the number or the sentence the AI attributed to it. If you cannot find it on the page, treat it as unsourced.
Should I delete every number I cannot verify?
Not necessarily. A number without a source can still be a useful estimate, as long as you label it as one and stop presenting it as a fact. Quotes are different. A quote you cannot find on the cited page should go, because it is a sentence someone wrote while pretending someone else said it.
Can I ask the AI to check its own sources?
You can ask it to search again, but the check that counts is the page you open yourself. A model asked to confirm its earlier answer will often confirm it. The authority is the page, not the model's description of the page.
Adrien Lebas
Founder of Banast, building it alone from Avignon. Writes about what it takes to make a living from what you build.