Attributing PR Campaigns to AI Answers

Consider a composite example. The retainer renews tomorrow. The agency sends a deck: eleven placements across outdoor and business press, two screenshots of a gear site’s homepage. The growth lead opens a chat log instead — one buyer question, “portable power station for a van conversion that runs a 500 W fridge overnight,” asked three times a week apart. Four brands come back each time; theirs appears once, as “a solid budget option.” Both documents are true. Only one answers whether the retainer worked.

PR earns GEO credit only when an answer moves inside a prompt set fixed before launch and holds against a control group; coverage proves the pitch landed, not that a model read it.

Four handoffs, and coverage measures one

A placement reaches an answer through four stages: published, independently crawled, rewritten by a third party, assembled into an answer. Coverage counts the first; the rest run on someone else’s clock.

A release on a syndication site sinks in days; the same study on your own domain, with a methodology page, data table, and update date, gets re-fetched for months. The third handoff is where value appears: a reporter or community writer takes your numbers and builds their own argument — what citable assets are for. The fourth is assembly: one article often contributes a clause of context, sometimes only that the brand exists.

A placement can surface a month later, and no single one supports “AI recommends us now.” Ask which handoff moved, and where next quarter’s budget goes.

Layer	What it measures	Data source	Common misuse
Reach	Whether the story ran	Placement links, media list	Reported as the result
Corpus	Whether it was crawled and rewritten	Citing domains, third-party articles, crawl logs	Syndication counted as independent
Answer	Whether it appears, is cited, is recommended	Repeated samples of fixed prompts	“It appeared,” no position or citation

Noise that looks like lift

Answers move on their own: another metro, account, or week, and the brand list changes. Much of a visible campaign gain is that churn.

Four treatments: repeat every prompt three times and keep only recurring phrasing; keep Black Friday weeks out of the baseline; hold ten category questions you never wrote for as a control; stagger paid social and creator drops, and say so when they cannot be separated.

Read the control first. If it rose as much as the test group, subtract it; if it fell while the test group rose, the signal is harder to explain away.

The brief that makes attribution possible

Attribution is an agreement made before launch, not arithmetic after it.

Step	Action	Owner	Passes when
1 Lock the set	20 buyer questions from tickets and site search, plus 10 controls	Growth, content	Each maps to a page or competitor
2 Store a baseline	Sample 14 days out, three runs each, save text and cited domains	Growth	One prompt’s spread is visible
3 Log the corpus	New citing domains, third-party articles on your data, crawl hits	Content, technical	Every domain has a first-seen date
4 Read	T+30, T+60, T+90 — same account, region, wording	Growth	Test and control in one table
5 Report	Three lines of conclusion, one page of detail, one next action	Lead	A decision in five minutes

Step 3 is the one teams skip and the one that explains causality: a citing domain usually appears before the answer changes, so skipping it credits the campaign’s last week for a rewrite that entered the corpus two weeks earlier. Log it as citation monitoring does, with one entity ID across links, mentions, and citations.

Example: one data campaign, four readings

Example — brand, figures, and campaign are invented; read the time series, not the numbers. A DTC brand runs a camping power survey with a methodology page, data table, and three expert commentaries: twenty test prompts, ten controls, four sample dates.

Metric (Example)	T-14	T+30	T+60	T+90
Test prompts naming the brand	3	5	9	11
Answers citing the brand’s domain	0	1	4	5
Third-party articles using the data	0	0	2	3
Control prompts naming the brand	2	2	3	2

Three readings. The 3-to-5 move at T+30 sits inside the control’s 2-to-2 spread: noise, not results. The inflection is T+60, as third-party articles go 0 to 2; that timing explains more than the count. T+90 flattens — corpus accumulation has a ceiling, so the next round needs a new data point, not another mailing.

Three lines, one page

Two audiences. For whoever funds the next round: which layer moved, whether it separates from control, where the budget goes. For the team: prompt text, sample time, answer text, cited domains, a verdict — new, gone, reworded. Keep mention, citation, and recommendation as separate counts, as the three layers of off-site authority do.

Do not report the test group alone; do not merge “mentioned” with “cited,” the difference between being in the room and being the source; do not claim a lift without the control’s same-period change; do not turn one sample into a finding.

Where this breaks

Event PR — funding, a recall, a lawsuit — drowns the signal; unify the message, do not attribute. Campaigns you cannot stagger cost more to separate than the answer is worth; report the total and label it inseparable. In-platform AI and agent shopping barely read press; they read listings, reviews, fulfillment, and feed data, so this covers off-site authority only. One extra mention written up as a lift discredits a small-sample report.

The most common misuse is choosing prompts after the fact, reporting only the questions that improved. That is evidence shopping, and a reviewer who asks for the excluded prompts will find it.

This month, before the next PR calendar is locked, put 20 buyer-voice questions, 10 controls, a 90-day window, and three read dates into the campaign brief, and archive one baseline sample before launch. Without it, the post-mortem is another argument. For the dashboard view, start with the measurement hub; for how off-site nodes corroborate one another, open the issue’s off-site trust network.