Earned media nobody pitched, tracked like coverage
AI assistants describe your brand to buyers thousands of times a day, using sources you did not choose and language you did not write. It behaves like earned media, it moves faster than a news cycle, and until recently there was no clipping service for it.
The PR playbook transfers better than most teams expect. There is a monitoring list, which here is a prompt set. There is a source to trace, which here is the page or profile the engine cited. There is a narrative to work, which here means changing what the cited sources say. What changes is the cadence and the measurement: answers update when sources are re-fetched, and the result is a mention rate and a tone reading per engine rather than a clip count.
Why PR teams track this
It is the first description many buyers see
For a growing share of research journeys, an AI summary is the first characterisation of your brand a buyer encounters, before your homepage, before any coverage you placed and before a sales conversation. It is assembled from whichever sources the engine trusts, in the engine's own words.
Sentiment is measurable per engine
How models talk about you differs by engine because they read different sources. One may lean on a review platform where recent feedback is critical; another on a trade publication that covered a launch well. Knowing which engine is negative and why points straight at the source to work on.
Crisis surfaces here early
A bad narrative appears in AI answers as soon as the sources it draws on shift, often before it reaches mainstream coverage and sometimes while it is still confined to a forum thread. A prompt set run on a schedule catches the change in tone within days.
The source is always named
Every cited answer carries the URL it drew on, which is something PR has never had from word of mouth. An inaccurate answer becomes a correction request to a specific publisher; a negative one becomes a specific page or profile to work on. The response plan writes itself from the citation.
Your AI brand monitoring workflow
Running AI answers as a monitored earned-media channel rather than an occasional spot-check.
- 1
Define the prompt set
Write the questions a journalist, an analyst and a buyer would actually ask about your brand and category: the category question, comparisons against named competitors, reputation questions ("is X reliable"), leadership questions, and any topic tied to a live issue. Keep it to a manageable size, fix it, and review it quarterly rather than editing between runs, because a changed prompt set makes the trend line meaningless.
- 2
Baseline mention rate, sentiment and sources
Run the set across the engines your audiences use and record, per engine and per prompt: whether you are mentioned, in what tone, alongside which competitors, and which sources are cited. Record the answer text itself, since the wording is what a buyer sees and what an executive will ask about.
- 3
Trace every answer you want to change back to its sources
For each negative, inaccurate or absent answer, list the cited sources and classify them: editorial, review platform, community thread, your own site, a partner. The narrative lives in those pages, not in the model, and the source type determines the response. A stale article needs a correction request; a critical review profile needs customer engagement; a forum thread needs genuine participation over time.
- 4
Work the sources by type
Pitch editorial with the story and evidence the publication needs to update or add coverage. Complete and refresh review profiles, and prompt satisfied customers at the right moment. Participate in community threads with people who know the product, never with a marketing account. Update your own pages only where the engine is already reading them.
- 5
Re-run the set on a schedule and report the movement
Weekly in an active category or during an incident, monthly in a stable one. Compare against the baseline per engine, and log which sources have changed or been added to the citations; a placement that reached the answer shows up as a new cited source before the tone shifts. Report mention rate as share of voice against named competitors, tone per engine, and the cited sources as the coverage actually shaping the answer.
PR and brand use cases
Share of voice against competitors
Who gets named when a buyer asks the category question, per engine, and how that changes over time. It is the AI equivalent of share of coverage.
Reputation monitoring
Catch a negative characterisation while it is still confined to a few sources, and find out which sources those are. The response is a targeted correction rather than a campaign, and the follow-up run confirms whether it worked.
Launch and announcement tracking
Whether the news actually reached the sources the models read, and whether the answers changed as a result. Coverage the engines never cite does not change what buyers are told, and this is how you find out which coverage did.
Executive and founder visibility
How AI describes your leadership by name, which buyers, candidates and journalists increasingly research directly. Include leadership prompts in the set and treat a stale or inaccurate bio as a source to update, not a curiosity.
Incident and crisis response
During an incident, run the reputation prompts daily and trace any shift in tone to its source. The engines repeat the sources they trust, so knowing which those are is the difference between a targeted response and a general statement nobody reads.
What differs per engine
Each engine reads a different mix of sources, so the narrative and the response differ by engine.
ChatGPT
ChatGPT's browsing answers cite editorial, review and brand sources inline and can reflect a corrected source within days of re-fetching it. Its non-browsing answers draw on training data and repeat an old narrative until the next model update, so a correction that has landed in the sources may not show here for a while. Track the two separately.
Perplexity
Perplexity is retrieval-first, community-heavy and fast to update. Reddit threads and forum posts surface in its citations more than in any other engine, which makes it the earliest warning of a shift in tone and the place where community work shows up first. It is the engine to run daily during an incident.
Google AI Overviews / AI Mode
AI Overviews draw on Google's index, so the sources tend to be pages that already rank for the reputation query: news coverage, review platforms and forum threads. A correction reaches this surface once Google has re-indexed the source. It is the engine most likely to reflect mainstream coverage and least likely to reflect a niche forum.
Gemini
Gemini grounds on Google Search and is used heavily in consumer contexts, so review platforms, shopping surfaces and mainstream coverage weigh more. For consumer brands it is often the engine where product-quality narratives show up most clearly.
The numbers to report
The PR reporting line for AI answers, and what each figure is safe to claim.
- Share of voice
- The share of runs on the category and comparison prompts where you are named, against each named competitor, per engine. It is the headline figure and the one most comparable to existing coverage metrics. Read it per engine, since a gain on one and a loss on another net to nothing blended.
- Sentiment by engine
- The tone in which each engine characterises you on the reputation prompts, with the specific criticisms listed. The criticisms matter more than the trend, because each traces to a source. A shift on one engine and not others points at a source only that engine reads.
- Cited-source mix
- Which sources each engine drew on for your prompts, by type, and how that changed since last period. A new source appearing in the citations is the earliest evidence a placement or correction has reached the answer, ahead of any movement in tone.
- Answer accuracy
- The share of prompts where the answer is factually correct about pricing, products, people and positioning, with the offending source noted for each miss. This is the figure that turns into a correction list.
Common mistakes and the fix
Mistake · Editing the prompt set between runs
Fix · Fix the set and review it quarterly. Any change to the prompts breaks the trend line, and the trend line is the entire value of the report. Add new prompts as a separate tracked group rather than replacing existing ones.
Mistake · Reading the model as the problem
Fix · The model repeats its sources. Trace every answer you want to change to the cited pages and work those; a general appeal to the engine changes nothing. The correction is always a page, a profile or a thread.
Mistake · Responding to Reddit with a brand account
Fix · Have someone who can speak for the company reply under their own name, correct the fact if it is wrong, and fix the product if it is right. A marketing tone or a fresh account makes the thread worse and keeps it cited longer.
Mistake · Blending engines into one score
Fix · Report per engine. Engines read different sources and move on different clocks, and a blended score hides a negative narrative on the engine your buyers use behind a positive one on an engine they do not.
A worked example: a consumer fintech app
A hypothetical consumer fintech app with a strong press profile finds that Perplexity and ChatGPT both describe it as reliable but with reported customer-support issues when asked whether it is trustworthy. The baseline on a thirty-prompt set shows the characterisation on most reputation prompts on Perplexity, some on browsing ChatGPT, and none on Google AI Overviews. Tracing the citations shows the source: two Reddit threads from the past quarter in which users report slow responses during a specific outage, and a review platform where the support score has dropped over the same period. No press has covered it.
The team classifies the sources as community threads needing genuine participation and a review profile needing customer engagement, with no editorial work required yet. The support lead posts in both threads under their own name, acknowledges the outage, explains what changed and invites anyone still affected to reach out. Support follows up with affected customers and prompts recent, resolved cases to leave reviews. The prompt set runs weekly, and the answer text for the reputation prompts goes into the weekly comms note so leadership can see the wording.
Within a few weeks Perplexity's answer shifts to noting the outage and the company's response, citing the updated thread; the review platform's support score recovers more slowly as reviews accumulate, and ChatGPT's browsing answers follow once the sources are re-fetched. At the quarterly report, share of voice on the category prompts is unchanged, tone on the reputation prompts has moved on two engines, the cited-source column shows exactly which pages changed, and the team has a time-to-correction figure per engine for the next incident.
Frequently asked questions
Is AI visibility a PR job or an SEO job?+
Both, split by where the work happens. Technical access and on-site structure sit with SEO. Which third-party sources describe you, and how, sits with PR, and that half tends to be the binding constraint once the technical work is done. Share the prompt set so both teams measure the same thing.
Can we get an AI model to correct something about our brand?+
Not directly, and be sceptical of anyone claiming otherwise. You change the sources the model reads. Retrieval-based answers update relatively quickly once the cited source changes and is re-fetched; anything grounded in training data will not move until the model itself is refreshed.
How often should we measure?+
Weekly for an active category or during a launch or incident; monthly is enough for a stable one. The value is in the trend, so keep the prompt set fixed between runs.
How does this fit our existing media monitoring?+
It sits alongside it. Traditional monitoring tells you what was published; this tells you what buyers are actually being told when they ask, which depends on what was published but is not the same thing. Together they show whether the coverage you placed reached the answer.
Which engines should a PR team track?+
Start with the ones your audiences use, which for most B2B brands means ChatGPT and Perplexity, with Google AI Overviews mattering wherever buyers still start on Google. Add Gemini for consumer categories. Measure across all once, then focus the effort where the audience is.
What counts as sentiment in an AI answer?+
The tone in which the engine characterises you: recommended, neutral, cautioned against, or described with specific criticisms. It is read from the answer text per prompt and per engine, and the useful part is the criticism itself, which almost always traces to a cited source you can act on.
What do we do when a Reddit thread is driving a negative answer?+
Read it first and check whether the criticism is fair. If it is, fix the product issue and let the thread age. If it is inaccurate, a genuine reply from someone who can speak for the company, correcting the fact without marketing, is the only response that works. Never astroturf.
How do we report this to leadership?+
As share of voice, tone per engine and the sources behind the answers, on a fixed prompt set, with the trend since last period and the answer text for the handful of prompts they care about. Keep it on one page next to the media monitoring summary.
Explore the features behind this solution
Go deeper
Take control of your AI narrative
See what every major engine says about your brand today, and which sources are saying it.