The page that measured a number and the page an AI engine credits for it are usually two different pages. You've seen the second half of that split: an AI Overview answers a question you hold real data on, and the citation underneath points at a roundup, an aggregator, someone's recap. Your page was crawlable, ranked, right there. The engine went past it, and it did so for a mechanical reason you can trace.
I watched this happen to our own studies often enough that we followed 1,006 blog citations to the very end of their reference trails. The average trail ran 1.20 hops. Only 17.2% ever reached a primary source. A trail that short ends fast, and it ends at whichever page happens to hold the number in a finished, readable sentence. That page collects the credit. Everything short of it gets treated as a relay.
Most advice about AI citations starts with quality: tighter answers, schema markup, more brand mentions. All of that runs after an earlier step almost nobody explains. Before an engine weighs how good your page is, it works out where your page sits in the chain of who said what, and that step is the one that keeps costing you the citation.
The Two Positions in a Citation Chain
Say one of your posts cites a 67% adoption figure and links the blog where you first read it. An answer engine picks up a related question, starts resolving that figure, follows your link, and finds the same 67% sitting on that blog in a tidy sentence with its own source line. The engine credits the blog. Your page and theirs just played the only two roles a citation chain has.
A hop is a page the trail passes through. It repeats a number it took from somewhere else, so the resolver keeps moving. A terminus is a page the trail stops at: the number lives there in finished form, nothing leads further, and the engine treats it as the origin.
Quality never gets a vote in this sorting. Your post can be the deeper analysis, the one a human actually bookmarks, and still get passed through, because the walk only asks one thing at each page: does this number lead somewhere else?
How engines choose and present answers in general is what AEO covers. The narrower question, which page in a chain collects the credit, is the one that stings, because the answer is so often a page that did none of the work.
How AI Resolves the Source of a Number
An engine resolving a claim behaves like a careful editor on a deadline. Who actually said this? Follow the link. Does the next page read like the source? Then stop.
Stopping comes quickly. The mean chain depth of 1.20 means the typical page sits a single link away from whatever supports it, and after that link the trail runs out. One click, one page, done.
Follow a single figure through the pipeline. A research firm measures something and publishes it in a report. A widely read blog covers the report, restates the figure in its own words, and links the firm. You quote the blog, because the blog is where you met the number.
Now the engine walks it. Your page: borrowed number, link out, keep moving. The blog: figure stated plainly, source line in place, everything a resolver needs, so it stops. Three pages carried the same number and the middle one took the credit, because it was the last page the walk could comfortably read.
That middle page is usually secondhand itself. In the same corpus, about two-thirds of SaaS blog claims are borrowed, meaning the measurement behind them was done by someone other than the author. A repackaged number in finished form, with a trail behind it too shallow to walk, looks exactly like an origin from where the engine sits.
The authority framing falls apart right here. The engine walks a trail, the trail runs out of road after roughly one step, and whichever page it's standing on when the road ends gets the credit. Domain strength never comes up.
You can run this same walk on your own pages. Citation Provenance, the claims layer inside LiquiChart, follows each claim's source link outward hop by hop and records where the trail ends and how deep it went.
Why Ranking Pages Go Uncited
You might hold the top organic spot for a query and still watch the AI Overview cite someone else. Plenty of people typing "why can my site rank but not be cited" are staring at exactly that screen, and the published data on engine behavior says the two outcomes have been drifting apart for a while.
Ahrefs found that 37.9% of AI Overview citations also appear in the top 10 organic results, down from roughly 76% a year earlier. Originality.AI puts the split near even: about half of AI Overview citations come from the top 10, and the other half from pages that sit nowhere in the top 100. Surfer measured from the citation side and found 67.82% of cited sources don't rank in the top 10 for the query they answer.
Ranking answers one question: how well a page places for a query. Citation answers a different one: which page a claim resolves to. Two separate contests, two separate winners, and a page can take the first every day of the week without ever taking the second.
That's why the Overview reaches past your ranked page for a research firm or a news desk. Those pages hold the end of a short, readable trail. Where those trails actually end is something you can count.
Where Citation Trails Actually End
We took every third-party citation in the corpus, all 3,299 of them, and followed each to its final destination. The endpoints sort into three buckets, and only one of the three is a place where a number originates.
5.2% of those citations reach a primary source. The other 94.8% end at a page with no followable link at all, or at a link that never gets to the origin. A dead end functions like a source in the engine's eyes: with no further hop available, the page sitting above the dead end becomes the terminus by default.
The trails that do carry links come with their own failure rate. Roughly one external link in five is dead, paywalled, or broken, so a chain that points somewhere often points at a wall, and the wall promotes whoever repeated the number last into the last readable stop.
If you publish original research, this is the case that stings. A blog restates your study. An engine resolves the claim. The blog's clean restatement is the reachable stop, so the credit lands one hop short of you, on a page whose entire contribution was retyping your finding with a working link. To watch a single claim fail this way up close, read what claim verification catches.
Reading the Citation Under an AI Answer
The unit underneath every one of these outcomes is the claim. In LiquiChart a claim is a verbatim span lifted from the page, the exact sentence as published, classified as a statistic, a source citation, a temporal reference, or a comparison. Register a borrowed statistic as a claim and its chain position stops being a hunch you form while squinting at an AI Overview.
Pull up an Overview for a question you hold first-party data on and check whose page sits in the citation line beneath it.
However the votes fall, the pattern behind them holds: the citation settles on whichever page presents a number in finished, reachable form, and that page is rarely the one that measured it.
Producing a number and getting cited for it are two separate events, and most pages only ever experience the first. The work happens on your page; the citation gets decided later, somewhere in the chain that grows out of it. As readers weigh in above, a pattern in who collects that citation will take shape, and it rarely lands back on whoever did the measuring.
How to Become the Page AI Cites
The mechanism points straight at the fix: build chains that end on you. A poll your own audience answered, a chart built from your own dataset, any number you measured rather than repeated, has no upstream. An engine resolving it lands on your page and finds nowhere further to go.
In practice that means running the poll yourself instead of quoting a vendor's survey, and charting your own usage data instead of citing someone's write-up of theirs. The figure starts on your page because it started with you.
The Claims layer registers a page's statistics as claims, traces each chain to its end, and grades the depth. You can run the single-number version of that check right now.
Paste a page you own and one statistic from it. The checker reads the live page and quotes back the sentence holding that number, so you can see whether the figure starts there or points away. A sentence that links out is telling you in advance where the citation will go.
There's a ceiling worth knowing about before you invest here. When Ahrefs traced ChatGPT's 1,000 most-cited pages, 67% fell into categories no outreach campaign can enter: Wikipedia, homepages, educational domains, app stores. Roughly a third of the most-cited slots are contestable at all, and inside that band the durable edge belongs to the page that held a number first. Engines reward the same distinctness in adjacent ways, covered in how to rank AI-generated content and Google's information gain score, and the page-level selection tactics live in how to get cited by AI search. Tactics decide which origin an engine picks. Owning the measurement decides whether you're an origin it can pick.
Own the Numbers That Matter
Citation position is structural. It was set by the links on your page the day you hit publish, and no ranking gain, redesign, or engine update will reassign it. Every borrowed statistic on a page you care about is a trail you built toward someone else's citation, and the day an engine walks that trail, the credit stops wherever the trail does.
So pick the page that earns you the most and find the one statistic on it you'd hate to lose credit for. Measure your own version. Watching where your trails end over time is its own discipline, covered in citation chain monitoring, but the first move takes one page and one number. Own at least one number on every page that matters, because until you do, the chain you published already decided the citation, and it decided against you.