Why does an AI assistant keep citing outdated facts from your product pages?

An AI assistant does not check whether your product page is current. It checks whether it is quotable. Those are different tests, and a page can pass the second while failing the first for months without anyone noticing. The failure mode is called fact drift: the fact you published in March is still structurally perfect in September, still well marked up, still easy to extract — and no longer true. The assistant keeps citing it anyway.

This matters more in an assistant-mediated channel than it did in search. A search result that shows a stale price sends the shopper to your page, where they see the correct one. An assistant answer that states a stale price ends the conversation with a wrong number attached to your brand. The merchant rarely sees it happen, because the answer is generated off-site and reported nowhere in your analytics.

## What fact drift actually looks like

Drift is not the same as a missing fact. A missing fact makes you invisible; a drifted fact makes you wrong. Four patterns cover most of what we see in scans:

Stale values. A price, lead time, or shipping window that changed in fulfilment but not in the page text. The most common variant is a promotional cutoff date that passed while the copy stayed up.

Contradicted values. Two pages state the same fact differently. The shipping policy says 7 to 10 days; the product page says 3 to 5. Each page is individually quotable. Together they are an unresolvable conflict, and assistants tend to resolve conflicts by excluding the merchant.

Orphaned values. A fact stated once, in a place that no longer exists or is no longer linked — a blog footnote from 2025 that says your warranty is two years, while the policy page now says one.

Ghost attributes. A spec that was true of an old SKU revision and survived a product refresh. Capacity, materials, compatibility, dimensions — anything where the physical product changed and the text did not.

## Why drift accumulates faster than it used to

Three structural reasons, none of them about carelessness:

1. Facts now live in more places. Price lives in the storefront, the marketplace listing, the feed, and the schema block. Changing one is a project; changing all four is a discipline. 2. Assistants cache. A remembered fact persists after the source changes, so the wrong version keeps circulating even after you fix the page. This is the pattern we covered in [why assistants reuse stale product facts](/en/blog/) — the correction has to outrun a memory. 3. Nobody is assigned to it. Drift is not a bug report. No customer complains that an assistant stated last quarter's price. It surfaces only if something is actively looking for it.

## A drift audit you can run this week

You do not need new tooling to start. The first pass is a text diff.

Step 1 — Pick your fifteen highest-value facts. Not pages: facts. Price ranges, lead times, warranty length, return window, materials, compatibility. The ones an assistant would have to state to answer a buying question about you.

Step 2 — Find every page where each fact appears. Use a site search for the number or phrase, not a page-by-page read. Include policy pages, blog posts, FAQ pages, and any landing pages from older campaigns.

Step 3 — Diff against the source of truth. The source of truth is not the marketing page. It is your fulfilment system, your current policy, your current feed. If the page disagrees with the system, the page is drifted.

Step 4 — Record which direction the conflict runs. Stale (old value, no competing version), contradicted (two live competing versions), orphaned (one live version, wrong authority). The fix differs by type.

Step 5 — Fix, then re-verify. Fixing is the easy half. Confirming the fix is the half that gets skipped, and it is the only half that tells you whether the conflict is actually gone.

## Fixing by drift type

Stale values need a single edit plus a date stamp. Where the fact is time-bound — a promotion, a seasonal window — state the window explicitly in the text rather than leaving the reader to infer it. A fact with a stated expiry cannot drift silently; it expires visibly.

Contradicted values need consolidation, not correction. Do not update one page to match the other and leave both standing. Two pages stating the same fact is a permanent liability, because both can drift again. Pick one canonical page and make the other reference it.

Orphaned values need removal or repointing. An old blog post stating a superseded warranty is worse than no post, because it is quotable and wrong. Either update it or delete it.

Ghost attributes need a refresh tied to your SKU revision process. If the physical product can change without the page text changing, drift is scheduled, not accidental.

## Structured data does not protect you

A well-formed JSON-LD block makes a fact easy to extract. It does not make it true. In fact, structured data that disagrees with the visible page text is one of the clearest failure signals an assistant can encounter — the markup says one number, the paragraph says another, and neither is marked as authoritative. Our earlier piece on [which JSON-LD schema types actually matter for AI product citations](/en/blog/which-jsonld-schema-types-matter-ai-product-citations-2026) makes the same point from the other direction: schema is a reinforcement layer, not a source of truth. If the text is drifted, schema amplifies the drift rather than correcting it.

The same logic applies to [FAQ schema](/en/blog/does-faq-schema-help-shopify-product-citations-2026). Marking up an answer that contradicts your return policy does not create a second opinion. It creates a contradiction with a machine-readable label on it.

## What to check weekly rather than yearly

If a full catalog audit is too heavy to run often, run a narrow one on the facts that change most:

• Price and promotion copy — weekly, because this is where drift originates fastest.
• Lead time and cutoffs — weekly during peak season, monthly otherwise.
• Return, warranty, and shipping policy pages — after every policy change, plus a monthly sweep.
• Product specs on refreshed SKUs — at every product refresh, as a gate rather than a follow-up.

The point of the cadence is not thoroughness. It is that drift found within a week is a copy edit, and drift found within a year is a citation history problem.

## FAQ

A content audit asks whether pages are accurate, complete, and up to date in a general sense. A drift audit asks a narrower question: which facts an assistant would state about us are stated differently in two live places, or stated in a place we no longer control. The scope is smaller and the fix list is more actionable.

Yes, if the page is still indexed and quotable. Traffic is not the criterion — extractability is. A low-traffic policy page is exactly the kind of page assistants read, because it is the authoritative-sounding source for a constraint.

Test the fact directly. Ask an assistant the question that would force it to state that fact — price, lead time, warranty — and see what it returns. If the answer differs from your current source of truth, you have found live drift. Run that check across your fifteen highest-value facts and you have a working drift baseline.

You can automate the detection far more cheaply than the fixing. Visora's free audit reports which facts are extractable from your pages and where inconsistencies show up across a catalog, so the weekly sweep becomes a review of a short list rather than a manual read. Start at [geovisora.com/audit](/audit) — the scan takes a few minutes, and the report names the specific pages rather than handing you a score.

Drift is not a content quality problem. It is a maintenance problem, and the merchants who handle it treat facts the way they treat inventory records: something with an owner, a source of truth, and a check. In an assistant-mediated channel, the version of your product page that matters is not the one you last edited. It is the one a machine can still state confidently — and that is a distinction you have to maintain on purpose. For more on the discipline behind it, our [FAQ page](/faq) covers how Visora tracks extractable facts over time.