Build a Citation Decay Alert System from Comparable Snapshots

A page with few AI citations in the current window may have lost them or may never have had any, and one status cannot say which. Compare two explicit, equal windows, then save the dated result so later runs remain auditable.

The GEO problem it solves

A low citation count looks like a drop, so the page goes into the refresh queue. But each page-presence status describes only the requested evidence window: WORKING, NEEDS_ATTENTION, NOT_CITED and DISCOVERED say how much evidence exists inside it, not why the count changed. A refresh booked on one window alone can spend writer time on a page nobody cited in either period.

One window answers what is cited now; two adjacent, equal windows answer what changed. That is also what separates this from measuring a campaign: no intervention or launch date is assumed. It runs on a schedule and asks whether ground you already held is eroding.

How the workflow breaks down

1. Map current citation health. get_page_presence returns monitored and discovered URLs with status, total citations, unique questions, average position and providers inside the requested window. Use the latest completed 30 days.

2. Reconstruct the preceding window. Call get_page_presence again with explicit start and end dates for the immediately preceding 30 days, then diff URL by URL. The tool does not return a time series, and the monitored-page registry is current, so save the resulting pair for auditability and to preserve any later configuration change.

3. Read the change the tool computes. get_source_trends compares source domains against the immediately preceding equal-length period and returns, per domain, current mentions, previous mentions and the percent change. Your own domain's row gives the site-level direction, and the third-party domains gaining fastest are displacement candidates. It ranks by current citations, so a collapsed domain can fall out of the rows, and the change is null when the previous period held no citations.

4. Read the surrounding evidence. get_source_urls filtered to your domain shows which pages are cited, get_source_domains the landscape around them, get_content_opportunities the prompts where competitors are mentioned and you are not. Give all three the same current-window dates and filters. A prominent competitor page is a displacement candidate, not a proven replacement.

What a run actually returns

A valid run returns one current and one previous row per normalized URL, each carrying the window dates, citation count and status. A page at one citation now and one before is stable; a page at one now and forty before is a decline. The current status alone cannot distinguish them.

The calls reconstruct evidence windows, but not historical monitored-page configuration. Keep the paired export when URLs change so a registry change is not misread as citation decay.

What to do with the output

Write only what the diff confirms. A URL enters the refresh queue when two snapshots at least 30 days apart show its citation count falling by half or more, or falling to zero from three or more. Order the queue by the citation count in the older snapshot, highest first: that is what you actually lost. Smaller drops, and anything backed by a single snapshot, go to the watch list and are not written this cycle.

Save every run as a dated pair of windows. The query can be repeated while the underlying evidence is retained, but the saved pair preserves the exact monitored set and decision record.

Before writing, check whether a queued URL was redirected or renamed since the earlier snapshot: a moved page loses citations without decaying. Preserve cited passages and stable URLs when you edit.

The Complete Prompt

I want to detect citation decay and build a content refresh plan. List the brands I monitor and, if there is more than one, ask which to use. Explain what each step reveals.

1. Define two adjacent completed 30-day windows in UTC: current and immediately preceding. Run get_page_presence once for each using explicit startDate and endDate, preserving any topic, tag, provider or funnel scope. Group each result by status and retain URL, citation count, unique questions, providers and exact dates. If both calls return no pages, tell me the brand has no monitored pages or no domain set, and stop.
2. Compare the two page-presence results URL by URL on the current monitored set. List count and status movements, and flag URLs present in only one result for registry or URL investigation rather than calling them decay. Ask for a saved snapshot only to detect monitoring-configuration changes, not because the evidence windows are unavailable.
3. Call get_source_trends for the current 30-day window with the same filters. Report my own domain's current mentions, previous mentions and change, then the third-party domains with the largest gains as displacement candidates. Read a null change as no citations in the previous period, not as stability. If my domain is absent from the rows, say so and use the page-window diff as the only trend evidence. Then call get_topics and repeat it on the two or three topic IDs I care about most.
4. Pull source URLs filtered to my domain, then source domains, then content opportunities, all with the current window's exact dates and filters. Name the prominent competitor sources and the prompts where I have room, labelling any replacement or causation as an unconfirmed hypothesis. An empty list is no evidence in that scope, not an improvement.
5. Build the queue. A URL is a confirmed decline only when the two snapshots are at least 30 days apart and its count fell by half or more, or to zero from three or more. Order those by the older snapshot's count, highest first; everything else goes to review candidates or the watch list. Give each entry its evidence and one specific update to test.
6. Ask which queued URLs were redirected, merged or renamed since the earlier snapshot, and drop those from confirmed declines: a moved page loses citations without decaying.