An agent is answering a question about your market right now. It lands on your post, takes the benchmark figure out of your third paragraph, and writes it into a report. That report feeds the next answer, and the next, in front of readers who'll never open your page.
AI agents and verifiable sources have collapsed into one problem. The agent has no way to know whether you measured that figure or carried it over from a survey you read once, so a number you produced and a number you relayed get the same treatment: acted on, at machine speed, with nobody in the path to catch the one that was already wrong.
Your page already gets picked. What decides the outcome now is whether the number on it survives a reader that never pauses to doubt.
How AI Agents Act on Your Numbers
An agent skips every step a careful reader takes. It treats each figure on the page as load-bearing, whether you measured it last quarter or lifted it from a deck in 2019, whether it still holds or moved a year ago. Then it acts.
For a decade the only consumer of your number was a person who could doubt it. Someone lands on the post, reads the sentence, decides whether to believe you. Worst case, one reader leaves with a wrong figure in their head. An agent changes the unit of damage. It takes the figure, builds on it, and repeats the result across a hundred answers before anyone notices the first one.
How badly this already goes has been measured. The Tow Center ran eight AI search tools through the same citation test, and the tools got more than 60% of the queries wrong, with two of them citing URLs that led to error pages more than half the time. That's the consumer your numbers now report to: confident enough to act, with a documented habit of walking dead links. And the moment your page relays a figure instead of originating one, the agent reads you as one more restating hop in a chain it's trying to reach the end of.
The Demand Side of Citation
Nearly everything written about content and AI answers a single question: how do you get cited. Answer engine optimization is a whole discipline built to make the engine pick your page and put your name on the answer. That's the supply side, and it deserves the work it gets.
The demand side gets almost none. Once the engine has your figure, is that figure safe to act on? The engine never asks. It can't tell whether you ran the survey or read about it in someone else's deck, and that gap shapes what happens next far more than which page collected the credit.
Put both sides on the same number. On the supply side you win: the engine names you, traffic ticks up, the dashboard looks healthy. On the demand side, that same engine has already carried your figure into answers you'll never see, and whether the number was yours to give never entered the calculation. A page can collect the citation and carry the consequence on the same number, and only one of those two shows up anywhere you can look.
Cite and act are different verbs. A citation wins attention. An action carries consequence, and consequence now travels at machine speed. So the question worth sitting with runs past whether you can trust the AI citations you read, to whether the citations on your own page would survive an agent that walks every one.
Why Verifiable Sources for AI Are Scarce
The distance between AI agents and verifiable sources can be measured directly. Take a borrowed statistic, follow its citation to the page it points at, and check whether that page actually states the figure. Do it once and you learn about one link. Do it across a whole corpus and you learn what the substrate is made of.
The trace covered 46 SaaS blogs and 3,299 borrowed claims, every one a figure some publisher had picked up from somewhere else, and we followed each citation until it ended. About seven in ten led nowhere a machine could check.
These are traced blog citations, human pages citing human pages, the layer the agents now read on top of. The State of Content Decay documents the same pattern at book length, and the full provenance study shows the tracing hop by hop. A name or a link beside a number makes it look sourced. Looking sourced and resolving to a source turn out to be different properties.
Walk enough of these chains and the dead ends sort into a few kinds. Some stop at an aggregator that re-cited the figure from a source it never names. Some reach a primary source that has since revised the number or deleted the page, leaving the figure floating free of anything that still states it. Some land on a page that never carried the claim at all. A thin slice resolves the way readers assume they all do: a page that measured the thing and still says so. That slice is where you want your numbers to live.
The scarcity is structural. The open web was written for humans who rarely click through, so a citation only had to look credible to someone skimming past it. Nobody graded whether the links resolved, because almost nobody followed them. An agent follows every one.
So ask the question from the agent's side. It just lifted a number off your page. What does it actually know about that number before it acts?
Every option names a real blind spot, and an agent carries all four at once, which is why a figure it can't place in a chain is a figure it acts on blind.
An agent reads the number exactly. What it cannot read is whether the number ends with you or passes through you, and that difference is the one that decides whether acting on it is safe. When the number only passes through you, there is nothing behind your page for the agent to check, and it acts anyway.
The Misattribution Problem Agents Inherit
Fabricated links, the failure everyone now pins on AI, barely exist in this corpus. Of every citation the trace rejected, only two across all 46 blogs pointed at invented pages.
The common failure is misattribution. Verify each link instead of trusting the one nearest the number, and about 35% of the citations that look linked point somewhere other than the claim beside them: a navigation item, a footer, a reference meant for a different paragraph. I checked a pile of these by hand, and the example that kept recurring was a store locator sitting next to a statistic it had nothing to do with. The link loaded fine. The claim beside it had come from somewhere else entirely.
This rot predates AI. The human-authored web has carried decorative citation all along, and to anything that judges provenance by proximity, decorative reads as sourced. A person skims past the mismatch. An agent follows the neighboring link, lands on a page that never made the claim, and proceeds as if it had.
Scale that across a substrate where about 35% of the linked citations miss their claim. An agent walking thousands of them inherits a misattribution rate the original authors never answered for, because nobody followed the links to find out. Those mistakes have sat in public for years. The agent is just the first reader diligent enough to inherit them.
Why Optimization Tactics Miss the Number
The reflex now is tactical: structured data, cleaner markup, an answer block inside the first 50 words, the moves that worked when the job was getting picked by search. None of them reach the number.
The information-agent era didn't create the fact-checking problem. It called in a debt publishing had been carrying behind blue links nobody clicked. Humans rarely walked a citation chain, so a borrowed figure with a plausible name beside it passed for a sourced one, and the borrowed-stat economy ran on that slack for years. An agent walks every link, and the slack is gone.
Formatting changes how a page presents a figure. Whether the figure is measured, current, or borrowed sits underneath the presentation, and that's the only layer that matters once an agent acts on it. Content provenance for AI agents lives in where the number actually came from. Mark up a relayed number perfectly and you have a perfectly marked-up relayed number. The move that changes the outcome is to own the number, to be the page where the chain ends rather than one it passes through.
Stronger models won't close the gap either. The most capable deep-research agents keep link validity above 94% while their factual accuracy sits between 39% and 77%: the links resolve, and the figures behind them have moved on. More retrieval can't repair a substrate that doesn't resolve.
What a Verifiable Source Means to an Agent
To an agent, a verifiable number is one whose origin it can resolve. Resolution is the whole test. It's how the agent decides what it can trust at all.
That's what a claim is: one figure pulled out of your prose and bound to the place it came from, an assertion the agent can locate rather than a loose string it repeats. LiquiChart's claim layer registers the figures that matter and tracks each one against its source.
Citation Provenance walks the chain behind a number, hop by hop, and records what sits at the end: a primary source you can stand behind, an aggregator that dead-ends, or a link that no longer states the figure at all.
From the outside you can only guess at another site's chain. The chains behind your own numbers are the ones you can read all the way down. Verifiable sources for AI come down to that reading running clean: a number an agent can follow to a page that still states it, which is the entire difference between a figure it can trust and a figure it merely repeats.
A terminus is only true on the day you check it. Sources revise, move, and disappear, and a number that resolved cleanly at publish can stop resolving six months later. Monitored Pages watch the external source behind a claim and flag the claim before an agent acts on the stale version.
You don't need any tooling to see the shape of this on one of your own numbers. Open a post you publish, find the statistic you'd least want to be wrong, and pull up the page you cited for it. Check whether that page still says what you cited it for.
When a Clean Source Goes Stale
There's a number on your site you're proud of. You measured it yourself, you've quoted it in talks, and it resolves cleanly to you, which is exactly why an agent trusts it most, repeats it most, and carries it furthest.
The day reality moves underneath it and the page stands still, that clean resolution becomes reach for a wrong number. The agent can't see the staleness. It keeps acting on the version it took, spreading your old figure across answers you'll never read, with your name attached to whatever it now gets wrong. Keeping a number true after publish is the line between being a source a machine can safely act on and being the one that taught it something false at scale.
Stale wears three disguises, and an agent sees through none of them. The source revises its figure upward while your page keeps the old one. The source deletes its page and your citation points at nothing. Or the number changes underneath a URL that still loads, so even a link check passes while the figure behind it has moved.
Somewhere in your published archive is the number you stopped checking. It's out working right now, at a speed and a reach you can't see, and whether it works for you or against you depends on when you last read the page behind it.