GEO-120·Issue 10·Authority & Off-site·120 / 132
Monitoring Citations in Third-Party Reviews
A model quotes one sentence from one review page, and that page expires. Log citations by page, sentence, and question — then fix the cheap causes first.
Issue focusMeasurable: AI Visibility, Attribution, and Experiments
Late September, a Berlin growth lead asks an assistant in German: which air purifier for a 45 m² flat with a cat and a toddler, under €400. Hers comes back second, backed by “replacement filters are cheap” — a line from a German test magazine’s spring buying guide. A year ago the same question pulled that magazine’s single-product test, which ranked her first. The publication did not change. The page did.
A review citation is not an asset you own; it is perishable inventory — the model quotes one sentence from one page, so log page, sentence, and question monthly, or you learn about the loss after the shortlist has moved.
The page is the unit, not the site
Off-site monitoring counts sites: this magazine covered us, that creator tested us. Answers resolve lower. Retrieval picks a URL, extraction keeps a span, synthesis writes around it, and any can flip while the domain stays put.
Page facts. A review showing €429 while your shop shows €399 gives the model a better number to read, or a reason to drop the review as stale.
Page structure. A spec table turned into an image, folded into an accordion, or stripped of units stops being extractable. Readable, no longer quotable.
Retrieval pool. A newer guide outranks an older test, an old page drops out of the candidate set, or phrasing shifts from “best air purifier” to “air purifier for a small flat with a cat.” Each swaps the cited object; your site never changed.
The unit is page × sentence × question: miss one dimension and a change cannot be explained. This is the drill-down under the citation layer of AI visibility metrics, which defines what counts; this one handles what follows.
A review-page monitoring table
Start with 8 to 12 review pages cited in the last three months, plus 3 to 5 that should have been cited.
| Field | Record | Why |
|---|---|---|
| Page | URL, language, market, type (test, guide, ranking, video) | Which questions it answers |
| Quoted sentence | The phrase pointing back to the page, verbatim | Only anchor comparable across re-tests |
| Question | ID and wording from a frozen prompt set | A moving denominator hides real moves |
| Position | Body citation, or a trailing source list | Read, not necessarily used |
| Role | Recommendation, neutral mention, counter-example | Still a citation, not a win |
| Co-mentioned | Other brands in the same answer | Who replaces you matters more |
The last two columns get cut first and matter most. A model can cite your page to explain “this unit runs loud” — the report says “cited once,” but the decision is wrong. Co-mentions point at the action: a rival’s new guide means a scenario test; a retailer’s ranking means a channel conversation.
Backlinks, mentions, and citations are three units; dedupe rules sit in citation monitoring.
Attribute before you intervene
The reflex when a citation drops is to email the editor. Rule out five causes first, from access to accuracy to prose.
| Symptom | Check first | Usual cause |
|---|---|---|
| Citation moves from A to B | Did B refresh price or comparison? | A newer page overrode the old claim |
| Citation holds, sentence turns negative | Context, test conditions, comments | New results or clustered complaints merged in |
| Everything drops at once | 404, paywall, login wall, region block | Access changed, not content |
| Only local-language prompts drop | Do local and English pages disagree? | Localization out of sync |
| Drops under one model or region | Is the sampling identical? | Different retrieval pool, not your fault |
Work access, then numbers, then prose. Reverse it and you rewrite a page whose only problem is a stale price.
One lost round is not a trend: the same question, another day, another account, returns another answer. Freeze the prompt set and sampling weekday, run three rounds monthly, and report movement only after two rounds agree. Co-mentions belong in the competitor benchmark table.
Example: one page replaced, one page repaired
Example — outlet, model, and figures are invented to show the record, not a real test.
A test magazine’s 2024 review listed €499, 5.8 kg, 24 dB on low. Through last year it was cited two or three times a month. From April it vanished from the same question set; the citation moved to the magazine’s new buying guide, which does not mention the brand.
Three findings. The review still showed €499, though the unit had been repriced to €449. The spec table had been rebuilt as an image. The guide added a “small flats, low ceilings” scenario the brand had never been tested for.
Three actions: operations sync master data and tell the editor; ask the outlet to restore the table as text; send a sample for the scenario.
Three months later the single test is back in body citations once; the guide still ignores the brand. The honest report line: one page repaired, one page still missing evidence.
Four interventions, cheapest first
The first two cost nothing; the last two cost budget and a season.
| Order | Action | Owner | Passes when |
|---|---|---|---|
| 1 Fix facts | Price, model, discontinued status, certification match master data | Operations, brand | Numbers match master data |
| 2 Request correction | Point at the error with public data; facts, not conclusions | Brand, PR | Correction live, correspondence filed |
| 3 Supply material | Testable conditions, comparison tables, local specs | Content | Cited unprompted, no paid placement |
| 4 Send a sample | A model or scenario the old page cannot answer | Marketing | New page enters the question set |
One line not to cross: you can correct facts; you cannot buy conclusions. Asking for a downside to be deleted or a rank specified buys one good sentence and devalues the source — models quote pages that read like selling points less often. What an outlet accepts depends on source type; where review corpora come from sorts the four kinds.
Where this breaks
A citation is not a recommendation: counter-examples and corrections land in the table too; the role column separates them.
Coverage is partial: paywalled reviews, video recommendations, podcasts. Say so when reporting “zero citations.”
If nobody has reviewed you independently, the table has no rows. Answers also move by region, account, and phrasing, so no screenshot is evidence.
Monitoring exists to make the next intervention less of a guess: which page expired, which sentence was misread, which scenario nobody tested. This line sits under authority and off-site, downstream of making reviews happen, upstream of turning citations into orders.
This month, take five real buyer questions and copy every review page cited in the answers into one table: URL, quoted sentence, and whether it recommended you or used you as a counter-example. Re-sample the same five next month on the same weekday. Explain only what changed; add metrics later.