How to Measure the Impact of Digital PR on AI-Generated Brand Mentions

Most PR teams can describe why AI search matters to earned media. Far fewer can show, on a given Monday, whether last quarter's campaign actually moved the needle inside ChatGPT, Perplexity, or Claude. Dashboards still default to impressions, domain authority, and referral traffic. None of those capture whether a model names the brand when a buyer asks a category question.

That gap is fixable, but only with a measurement system built for generative answers rather than ranked blue links. The workflow below is the one comms operators can run with a spreadsheet, a prompt library, and a disciplined log. It covers baselining, logging, share-of-model math, and attribution to specific placements, so a campaign's lift stops being an anecdote and starts being a number.

Start With a Baseline Prompt Set

Measurement collapses without a stable question set. If the prompts drift, the trend line is noise. Build a library of 60 to 150 prompts before any campaign runs, and freeze it. Treat the library as the control.

Three categories cover most PR goals. Branded prompts ("What is [Brand]?", "Is [Brand] legitimate?") confirm entity recognition. Category prompts ("best SOC 2 automation tools", "top venture firms investing in climate") measure recommendation share. Comparison prompts ("[Brand] vs [Competitor]") capture how the model frames the brand against a known rival. Keep the ratio roughly 20/60/20 so category coverage dominates without starving the other two.

Each prompt needs metadata: the intent, the target persona, the geography, and the business outcome it maps to. A prompt about "best enterprise MDR vendors" is only useful if it ladders to a real deal the brand would want to be shortlisted for. This is also where digital PR services and the campaign calendar should align, so the stories a team pitches match the prompts the team tracks.

Run each prompt multiple times per model. SparkToro's testing found less than a 1-in-100 chance that ChatGPT or Google's AI returns the same brand list twice, so a single query tells you almost nothing. Ten runs per prompt per model, every two weeks, is a defensible floor. That cadence also matches how fast content ages in AI answers: the May 2026 Generative Pulse report found that more than half of journalism citations in AI answers come from articles published within the past 12 months, with volume dropping sharply after six months.

Build the Citation Log That Powers Everything Else

The log is the primary asset. Every answer the system records gets one row, and the row is granular enough to answer any downstream question without re-running the query.

Required columns, at minimum: date, model, model version if exposed, prompt ID, run number, brand mentioned (yes/no), mention position (ordinal), citation present (yes/no), cited URL, cited publication, competitors mentioned, sentiment (positive/neutral/negative), and the full response text. Store the raw text. Six weeks later, a reviewer will want to re-score sentiment or search for a phrase that was not in the original schema.

The log is also where platform differences become legible. Claude mentions brands in 97.3% of responses, ChatGPT in 73.6%, and Perplexity between 40% and 48.5%, with Perplexity carrying external links in over 77% of responses versus ChatGPT at 31%. A raw aggregate across models hides that mix. Reporting per model, with per-model baselines, is the only honest way to read the trend. Siteimprove's framing of four core answer-engine metrics — share of voice, citation rate, brand sentiment, and prompt coverage — is a useful spine for the log, and none of those columns come out of GA4 or Search Console.

1. 1Define schema: date, model, prompt ID, run #, mention, position, citation, URL, sentiment, raw text
2. 2Freeze the baseline prompt library and assign stable prompt IDs
3. 3Run each prompt 10x per model on a two-week cadence
4. 4Write every response — one row per run — into the log, raw text included
5. 5Score mentions, position, and sentiment against the schema
6. 6Annotate platform events (model updates, interface changes) in a separate column

Calculate Share of Model Honestly

Share of model is a frequency, not a ranking. The core formula is simple: brand mentions divided by total brand mentions across the tracked response set, multiplied by 100. Run 100 prompts across four models at 10 iterations each, and the denominator is 4,000 responses. If the brand appears in 1,200 and all competitor brands together account for 6,000 mentions, raw inclusion is 30% and share of model is 20%.

Two refinements matter. First, report inclusion rate and share of model as separate numbers. Inclusion tells you whether the brand is in the conversation at all; share tells you how much of the oxygen it is taking from rivals. Blending them hides which lever moved. Second, weight by position when the model returns an ordered list. A brand named first in a top-five list is doing different work than one named fifth, and buyers do not read ranked answers evenly.

Set expectations against category reality. Semrush and Kevin Indig's analysis of 1,094 US categories tracked monthly in ChatGPT from January through June 2026 found that only 15% of AI search categories have a clear brand winner, while in the other 85% no single brand shows up consistently. Translation: in most categories a share of model above 15–20% is already a leading position, and chasing 50% is a misread of how fragmented these answers actually are.

Attribute Lift to Specific Placements

Attribution is where the system earns its budget. The question is not "did share of model go up?" but "which placement moved it, and on which model?" The log makes this answerable because every response carries its cited URLs. A rising share with no new cited domains means the model's priors shifted. A rising share with new cited URLs means a specific piece of coverage entered the index.

Run the attribution loop on a two-week clock. For every new campaign placement, record the publication date, URL, and the prompts it should plausibly influence. Watch the citation log for appearances of that URL (or its syndications) over the following 8 to 12 weeks. When the URL starts appearing as a cited source and the brand's inclusion rate on the mapped prompts rises in the same window, that is the attribution signal. It is probabilistic, not courtroom-grade, but it is defensible.

Third-party coverage carries disproportionate weight here. Agility PR research cited by MadX found a median 239% lift in AI search visibility when the same content ran through third-party outlets rather than sitting on the brand's own site. That is why the attribution map should privilege earned placements over owned content, and why the difference between mentions and momentum matters: a single tier-one feature that gets quoted inside answers for six months beats a flurry of posts that never clear the citation threshold.

Validate With Downstream Signals

Share of model is the leading indicator. Referral behavior is the lagging one, and both need to be in the same report so nobody confuses visibility with revenue. Previsible's July 2026 AI Traffic Study of 6.77 million sessions across 166 GA4 properties found that ChatGPT drives 92.4% of all trackable LLM referral traffic, with monthly sessions growing 12.8x over the 19-month window. That same analysis showed ChatGPT referrals converting at 15.9% and Perplexity at 10.5%, compared with 1.76% for Google organic on the same site.

Set up GA4 to segment LLM referrers as a dedicated channel group. Watch for coupled movements: a share-of-model lift on a cluster of prompts, followed by a referral and conversion lift from the matching model. When OpenAI shipped its May 7, 2026 interface update, Similarweb data showed total ChatGPT referrals rising 157.7% week over week and homepage referrals rising 354.7%. That is the kind of platform change that will distort any trend line, and the log needs an annotation column so platform events do not get mistaken for campaign effects.

For teams building the full pipeline from pitch to measurement, pair this with the mechanics in getting cited by ChatGPT, Perplexity, and Claude and the longer argument in the numbers behind digital PR in AI search. The measurement workflow is only useful if the campaigns feeding it are built to be citable in the first place.

Operating Cadence and Who Owns What

Weekly: run the prompt set, update the log, flag anomalies. Biweekly: recalculate share of model per model, update the attribution map, mark new placements. Monthly: a one-page report that pairs share-of-model movement, newly cited URLs, and LLM referral and conversion data from GA4. Quarterly: audit the prompt library, retire prompts that no longer map to pipeline, add new ones for emerging buyer questions.

Ownership is simple. One analyst owns the log and the math. The PR lead owns the attribution map and the placement-to-prompt mapping. The head of comms owns the monthly narrative that connects the three. Teams that split those duties across four people usually end up with a dashboard nobody trusts. Teams exploring external help can see how this fits into a wider engagement in how campaigns actually work.

None of this proves a direct causal line from a single pitch to a single AI citation. It does not need to. What it produces is a repeatable, platform-aware record of where the brand sits in generative answers, which pieces of coverage are doing the work, and which prompts still have no answer. That is the measurement system AI-era PR has been missing, and it is the one that lets comms stop defending its budget with press-hit screenshots and start defending it with share of model.