Audit Source Concentration Across Your Monitored AI Answers

If a handful of URLs carry most of the citations in your category, one editorial change on one page reshapes the landscape you depend on. This workflow inventories the cited domains, states how concentrated the mix is, and ends on a tiered backlog.

The GEO problem it solves

Source work usually gets handled as link building: chase authority scores, buy the biggest publishers. Authority lists describe the open web in general, not the domains an engine quoted in your category.

The citation evidence replaces the ranking question with a concentration question: what was cited, how often, and how much sits on very few domains. That is a risk lens, not a causal model: the observed mix cannot prove a new source raises visibility, or predict what losing one would do.

How the workflow breaks down

1. Inventory the cited domains. get_source_domains ranks domains with citation evidence in the requested scope by mention count, with source type, average position and backlink flag. It defaults to 30 days and accepts custom dates plus provider, topic, tag and funnel filters. The 100-row cap still makes it a ranked inventory, never the denominator.

2. Get a real denominator. get_source_trends counts citations inside a window (30 days by default, 365 at most) and returns total citations, unique sources, the full source-type mix with shares, and up to 10 leading domains with a previous-window count and percent change.

3. State the concentration, do not grade it. Divide the largest returned domain's mentions, then the summed mentions of the leading rows, by total citations, and print both beside the window and row count. No published figure says what those shares should be, so no band applies. Percent change is null when the previous window was empty.

4. Open the heaviest domains. get_source_profile takes a domain name or its ID and returns, for the window, mention and question counts, average position, providers, up to 20 top URLs and up to 10 co-cited SELF or DIRECT competitors. Domain diversity often hides URL-level concentration.

5. Add the competitor overlap. get_competitors supplies the competitor domain IDs, then get_competitor_position returns share of voice, coverage and the 10 domains citing that competitor most. Ten rows sample a portfolio, so the overlap stays directional.

What a run actually returns

A valid run returns a capped ranking beside an uncapped citation total for the same evidence window. The leading rows can account for only part of that total, and a high-volume domain can still have a weaker average position than a smaller source.

Those two fields answer different questions. Mentions show concentration; average position shows where the source tends to appear when cited. Work the rows where the observed concentration, position and competitor overlap all support an investigation, without turning any one field into a domain-authority score.

What to do with the output

Rank the backlog in three tiers. Tier one, leading domains whose percent change is negative: a dependency shrinking on its own is the one case where doing nothing already moves your mix. Tier two, domains present both in your inventory and in a competitor's citing set. Tier three, everything else, including source types with no returned domain: investigation candidates, not confirmed gaps. Order by mentions inside a tier.

If total citations comes back at zero there is nothing to concentrate: widen the window to 365 days, and if it stays at zero, report that no citations were recorded and stop rather than falling back on the undated inventory. Otherwise save the window, both shares and the date, since two snapshots built on different windows produce movement that exists only in the arithmetic.

The Complete Prompt

I want to audit source concentration across my monitored AI answers. List the brands I monitor and ask me which to use if there is more than one. Explain what each step shows before moving on.

1. Pull get_source_domains for the last 30 days, grouped by source type and ranked by mention count. Preserve any provider, topic, tag or funnel scope for the rest of the workflow. The list is capped at 100, so call it a ranking and not a denominator.
2. Pull source trends over the same 30 days with a limit of 10 leading domains, and report total citations, unique sources and the source-type mix with each share. If total citations is zero, retry both source trends and source domains once over 365 days with the same filters; if it is still zero, report that no citations were recorded and stop.
3. Divide the largest returned domain's mentions, then the summed mentions of all returned leading rows, by total citations. Show both shares with the window and row count. Do not band them: no benchmark exists for what they should be. Treat a null percent change as an empty previous window rather than growth.
4. Run the source profile on the five heaviest domains over the same window you settled on in step 2: mentions, questions, average position, providers, top URLs and co-cited competitors. It is a window aggregate, so do not read a trend from it.
5. Pull my competitors for their IDs using the same evidence window and filters, run competitor position on the top three by mentions with that identical scope, and compare the domains citing them against my inventory. Only ten source domains come back per competitor, so label the comparison directional.
6. Build the backlog in three tiers: leading domains with a negative percent change, then domains present in both my inventory and a competitor's citing set, then everything else including source types with no returned domain. Order by mentions inside each tier, and give every line its observation, tactic, owner and the number to re-check next snapshot.

Keep category-level citation data separate from claims about my brand, and never state that a change in the mix will produce a citation or a ranking outcome.