When Reputable Sources Disagree: The Answer-Variance Method

Every topic worth writing about has questions where the good sources do not agree. The instinct is to pick a side quietly and move on, which produces a page that is confidently wrong half the time and unhelpful the rest of it. There is a better output: publish the reconciliation, and decide the contested answers before anyone starts drafting.

This method has a name in our builds — the answer-variance registry — and it is a research artifact first and a content plan second. What follows is how to build one, what a reconciliation page owes the reader, and the honest limits of the whole idea.

How do you find the questions where sources actually disagree?

By collecting them deliberately, during research, before any page exists. Work through the questions your readers ask and flag every one where two reputable sources give different answers, then keep the flagged list in a single file. On our auto-finance rebuild that list ran to 20 questions; on our leasing build it ran to 18 [our data].

Two sources feed the list. The first is reading — following the same question across the primary sources in your field and noticing where the numbers stop matching. The second is your own query data, because the questions people ask repeatedly are the ones worth resolving; the mechanics of pulling those are in mining Search Console for AI-era queries.

Keep the registry small and real. A list of 20 genuinely contested questions is more useful than a list of 200 where most entries are one source paraphrasing another. Tracing each claim back to its origin is what separates the two, and it is the same discipline documented in how we fact-audit at scale.

Why do reputable sources contradict each other?

Usually because they measured different things and used the same word for both. Genuine contradictions — where both statements cannot be true — are the label to reach for last, after every measurement explanation has been ruled out. We have not classified our own registries by conflict type, so we cannot tell you what share falls into each row below; treating every disagreement as a contradiction is simply the fastest way to write an unfair page.

Classifying the conflict is not a formality. It decides the shape of the page — a definitions conflict needs a glossary paragraph, a period conflict needs a dated table, and a genuine contradiction needs a methodology argument you had better be able to defend.

What does a reconciliation actually look like?

Both numbers, both publishers, both periods, and a sentence explaining what each one measured. This library runs one live example of its own: Pew's browsing study found users clicked a cited source in about 1% of visits that included an AI summary (March 2025 data), while Seer Interactive measured sites cited inside an AI Overview earning roughly 35% higher organic CTR than uncited sites (2025).

Read across the rows and the conflict dissolves: Pew counted a behavior almost nobody performs, and Seer compared outcomes between two groups of sites. Both are true, and the reconciliation — what an AI citation is actually worth once you separate the pipe from the marker — is worked through in full in are AI citations worth anything.

Two rules keep a reconciliation honest. Never quietly average two figures from different periods into one number, and never present the source you prefer as "the data" while relegating the other to a footnote. Google's people-first guidance frames the bar as demonstrated reliability and depth (Google, helpful content guidance); a page that hides the inconvenient source fails that bar even when its conclusion is right.

Does a reconciling page actually earn citations?

Nobody knows, including us. The hypothesis behind the method — that a page resolving a documented contradiction is more likely to be selected as a source — is plausible-but-unproven, and we label it that way every time it appears. No study on our closed source list measures it, and our own logs cannot isolate it from everything else that differs between our pages.

What we can say is narrower and still worth something. A reconciliation page tends to be the only page on its question that carries both figures with their periods attached, which makes it more useful to a reader and more quotable as a passage. Usefulness is the part we control. Selection is not, and no one should tell you otherwise.

What happens if you skip the registry?

Your own writers resolve the contested questions independently, and reasonable people resolve ambiguous questions differently. On a day-one adversarial review of a 67-page build, 4 pages answered a single contested question 3 different ways — on the very build whose strategy was correcting other people's contradictions [our data]. The full incident, including why per-file gates cannot catch it, is in four pages, three answers.

The fix is ordering, not detection. Record each resolved answer, its reasoning, and its sources in a canonical-stances file that every writer reads before drafting. Writers then inherit a resolved question instead of resolving it themselves, which costs an hour before the fleet starts and cannot be bought back afterward.

When is "nobody knows" the honest answer?

When the sources disagree and none of them has a methodology strong enough to settle it. That page is still worth publishing — as a page that states the question, shows what each source found, names what would resolve it, and stops.

Against our own interest: this method is slower than picking a side, and it produces pages that end in qualified answers rather than confident ones. Qualified answers convert worse. We accept that trade because a reference library sells nothing but the reliability of its answers, and the alternative — sounding certain about a genuinely contested question — is the failure mode this library was built to avoid. The wider system it belongs to is documented in our GEO guide.