In a recent note, I broke down what actually gets a page cited by AI systems. The one finding with real experimental evidence behind it: specificity wins. The GEO study presented at KDD 2024 tested nine content modifications across roughly 10,000 queries, and the clear winners were adding statistics, adding quotations, and citing sources, each boosting visibility in AI answers by 30 to 40%. Pages that commit to concrete numbers get pulled into answers. Pages that hedge do not.
The takeaway everyone is going to run with is obvious: add statistics to your content.
That takeaway is half right, and the wrong half will hurt you. So this week I’m giving you the tool I use to get the right half: a free Claude skill (and a GPT version) called Receipts. It finds statistics that genuinely strengthen your content, verifies every one against its primary source, and refuses to hand you anything it couldn’t confirm. Download links are below, but first let me convince you the verification part is the whole point.
The Statistics Supply Chain Is Garbage
Here’s the problem with “just add statistics.” Most of the statistics you’ll find are broken in ways that aren’t obvious until someone checks.
The web is full of zombie statistics: figures that get cited everywhere and trace back to nothing. Someone published a number years ago with no methodology, it got repeated, the repetitions got cited, and now fifty pages “confirm” it by pointing at each other. That’s circular reporting, and it’s the most dangerous failure mode in fact-checking, because volume of repetition looks exactly like corroboration and isn’t. Some of these figures circulate for decades. You’ve probably quoted one. So have I.
Add to that the misattributed figures (the study exists, but it never says what the blog post claims it says), the mutated ones (a range becomes a point estimate, millions become billions, a 2019 figure gets a 2026 dateline), and the “50 Statistics About X” listicles that launder all of the above into tidy bullet points, and the picture is clear. The statistics supply chain most content teams draw from is contaminated, and drawing from it confidently is how bad numbers spread.
And the stakes just went up. The same AI systems that reward statistics also repeat them. When your fabricated or mutated figure gets cited in an AI answer, it doesn’t just sit on your page anymore. It propagates at machine speed, with your site as the source. An unverified statistic has always been a trust liability. Now it’s a trust liability with distribution.
There’s a second, quieter problem too: LLMs make statistics up. Ask a model for “statistics about X” and it will cheerfully produce plausible numbers with plausible attributions, some real, some assembled from vibes. If your plan is “have AI add stats to my content,” you’re automating the contamination.
So the play is not “add statistics.” The play is “add verified statistics.” That’s slower, which is exactly why almost nobody does it, which is exactly why it’s worth doing.
What the Claude Receipts Skill (and GPT) Does
Receipts is a skill I built for my own content workflow. It gets its name because not only does it find statistics, but it also brings “the receipts” for verification.
It does two jobs: finds statistics worth adding, and verifies them to a standard that would survive a hostile fact-checker. It runs against whatever you give it: a brief, an outline, a draft, a topic, or a URL. Point it at an existing page and it will also audit the statistics already on it, reporting each one as confirmed, misattributed, outdated, misrepresented, or unverifiable.
The design has one governing principle: an unverified statistic is worse than no statistic at all. Everything else follows from that.
It decides whether a statistic is warranted before it searches for anything. This is the part that separates it from every “add stats for GEO” tool you’ll see. A tool built to find statistics will always find statistics, and without restraint you end up with a number bolted onto every section, decorating rather than informing.
Receipts walks your content heading by heading and applies a removal test to every candidate placement: if this figure were removed, what would the reader actually lose? Concrete answer, it’s warranted. Vague answer (“it adds credibility”), it’s not.
There’s no quota and no ceiling. A data-heavy comparison page might warrant a dozen figures. A how-to guide might warrant zero, and “no statistics warranted” is a valid, successful output. Most content supports far fewer statistics than people bolt onto it.
Nothing comes from the model’s memory. Ever. The skill’s hard rule is that a statistic doesn’t exist until a real search returned it, the source was actually fetched, and the figure was read in the source’s own text. Familiarity is treated as a danger signal, not verification, because the figures that feel most obviously true are the ones most likely to trace back to nothing.
It’s also forbidden from constructing URLs, and from rounding, converting, or rephrasing figures. What the source says is what you get.
Every figure gets traced to its primary source, and the source gets opened and read. Blog cites article, article cites report, report cites study: the study is the source, and the skill follows the chain until it gets there. Then it fetches the actual document and confirms the figure appears in it, exactly. Search snippets don’t count. If the primary source is paywalled or unreachable, the figure doesn’t publish, no matter how promising it looked. A statistic you couldn’t open and read is a rumor.
It hunts circular reporting specifically. When multiple sources repeat a figure, it traces each one back independently. If they all converge on a single unsourced origin, that’s one source, not many, and if the origin doesn’t hold up, the figure gets rejected entirely, regardless of how universally it’s accepted.
It audits context, not just accuracy. A real figure can still be misleading: a study of 200 enterprise buyers in Germany presented as “customers want X,” a percentage-point change reported as a percentage, an outlier pulled from a study whose actual conclusion points the other way. The skill checks population, sample size, timeframe, methodology, and funding, and applies a simple test: would the researchers who produced this figure object to how you’re using it?
It checks whether something newer exists, then re-verifies the newer figure if so. And then, because verification fatigue is real, every surviving statistic gets a second, fully independent recheck before output: source re-fetched, figure re-located, character-for-character match confirmed.
What comes out the other end: each cleared statistic attached to the specific heading where it belongs, with the exact figure, publisher, study title, date, a live link to the primary source, a confidence tier (A for independently corroborated, B for single-source with visible attribution), and suggested phrasing that doesn’t overstate the finding.
Plus a rejection log showing every figure that was considered and discarded, and why. The rejection log is part of the deliverable on purpose. A run that returns one verified figure and eleven rejections is a better run than one that returns twelve unverified ones.
The Double Play
One more reason this is worth the effort, beyond the citation evidence.
Verified statistics content has always been one of the best linkable assets you can build. Writers and journalists constantly need a credible figure to cite, and a page that provides one, with a clean primary-source trail, earns links passively for years.
I covered statistics pages in my link building note for exactly that reason.
Now the same asset works a second channel. The figures that earn you links from human writers are the same figures that get you pulled into AI answers. One piece of verified work, two distribution systems, and both of them compound. Meanwhile the sites pumping out unverified stat listicles are building on sand in both channels at once.
Get the Skill
Claude version: SKILL DOWNLOAD LINK
Install it in Claude: Settings, then Capabilities, then Skills, then upload the file and toggle it on. Web search must be enabled, because the skill’s verification gates require actually fetching sources. If it can’t fetch, it refuses to output figures rather than downgrade to guessing, which is the correct behavior and the entire point.
ChatGPT version: GPT LINK
The GPT works the same way using ChatGPT’s browsing: it must open a source and read the figure in it before citing it, and if it can’t open the source, the figure doesn’t publish.
Use either one the same way. Hand it your draft, brief, or URL and say “run receipts” or “find statistics for this.” Or point it at a published page and say “verify the stats on this page” to audit what’s already there. That second use is worth running on your most important pages even if you never add a single new figure, because you almost certainly have a zombie statistic living on your site right now.
The Honest Caveats
Consistent with the citations note, two things to keep straight.
Statistics raise your odds of citation. They don’t guarantee it. The cross-engine research shows the tactical tricks don’t transfer reliably between systems, and citation patterns shift when retrieval systems change. What holds up is the underlying quality: a page with concrete, verifiable, well-sourced claims is easier for any system to ground an answer in, whichever system is doing the grounding that month.
And restraint is part of the strategy, not a limitation on it. The goal was never a number in every paragraph. It’s the right figures, verified, in the places where they genuinely resolve a claim. That’s what serves the reader, and as always, that’s the version that survives.
