Earned Media Is How AI Describes Your Brand

Earned media is how AI describes your brand. When a model answers category or reputation questions, it is often compressing third-party sources into a short description, not reading your homepage the way a salesperson would. Delve (delve.news) helps communications teams see what that record is made of upstream: coverage quality, message pull-through, and narrative framing, so you know what is entering the AI-era description of your brand before you chase citation scoreboards.

This page is about the inputs to how brands get described. It is not a guide to counting ChatGPT citations, and it does not treat share of model as a Delve product claim. For trust limits on compressing coverage with AI, see can AI coverage summaries be trusted. For framing over time, see what narrative tracking is. For share of voice inside the conversations you meant to own, see the new share of voice.

What you need to know — TL;DR

Third-party sources feed the description

Where do AI answers about brands mostly come from?

Muck Rack's What Is AI Reading? study has run three editions since July 2025. The May 2026 edition analyzed more than 25 million links cited by ChatGPT, Claude, and Gemini across 17 industries, and found earned media accounted for 84% of citations while paid and advertorial content accounted for 0.3%. Muck Rack reports earned media ranging from 82% to 89% across all three editions, and the May report shows journalism holding between 20% and 30% of citations across every study it has run. One caveat worth carrying: Muck Rack's earned bucket is broader than journalist coverage. It also includes academic, government, encyclopedic, social, and third-party corporate sources, with journalism itself around a quarter of citations. The takeaway holds either way. Non-paid third-party text dominates, so substance matters.

Consistency is the finding, not the percentage

How much should you trust a number like that?

The useful signal is that three editions over ten months, on samples from one million links to twenty-five million, keep landing in the same band. One snapshot could be a model update. A held pattern is closer to how these systems source information. Note also that Muck Rack sells a product in the AI-visibility category, so read it as vendor research with a large public sample.

Citation share is not the same as being described well

If you appear in answers, are you winning?

Not necessarily. Being cited for the wrong frame, a lagging product, regulatory doubt, or a competitor's category language, is visibility without control. The useful question is what substance of coverage exists for models to compress.

Volume is a weak proxy for the record

Why doesn't more coverage automatically help?

Clip count encodes neither outlet weight, message accuracy, nor narrative direction. Fifty off-message mentions can leave a worse source record than five on-message pieces in outlets that matter.

Measure the three inputs you can actually run

What should teams track week to week?

Coverage quality, meaning which outlets with what readership and relevance. Message pull-through, meaning whether configured claims appeared. Narrative and theme framing, meaning how topics are characterized and whether you are present in the themes strategy named.

Downstream AI-answer tracking is a different tool category

Should Delve replace a GEO or AI-visibility platform?

No. Those score presence inside ChatGPT, Perplexity, Gemini, or AI Overviews, which is a downstream question. Delve answers the upstream one: what earned evidence exists for those systems to draw on, and whether it matches the story you intended.

The shift comms leaders are already feeling

For thirty years, earned media's job was mostly human: reach journalists' readers, shape stakeholder perception, leave a trail of third-party credibility. That job has not gone away. What changed is that the same trail is now a primary input to systems that never open your media room.

Buyers, candidates, investors, and journalists ask models category and reputation questions. The answer is a compression of sources the system retrieved. When those sources are overwhelmingly editorial and other non-paid coverage, communications is no longer only awareness work. It is stewardship of the public text that becomes the brand's machine-readable description.

An earlier edition of the same research is worth sitting with. Muck Rack compared the journalists PR teams pitch most often against the journalists AI engines cite most often for those brands, and found an average overlap of 2%. Whatever you make of the precise figure, the gap it describes is the practical version of this whole argument: the coverage shaping the description may not be the coverage your program is aimed at.

That does not mean every clip is equally useful, or that appearing once in an AI answer proves reputation. It means celebrating volume alone is more expensive now than it used to be.

What "the AI record" means, and what it does not

Use "AI record" as a plain-language label for the pool of third-party coverage and other public text that models can draw on when they describe you. Not as a claim that one permanent database exists, and not as a claim that Delve stores model training corpora.

Three distinctions keep the language honest:

Citation is not verification. Visibility is not credibility. A summary of coverage is not the coverage. That last one is the same discipline we applied to AI summarization generally: orientation is useful, accountability needs structure, provenance, and judgment. See can AI coverage summaries be trusted.

Why quality, message, and narrative decide what enters

If models compress third-party text, the attributes of that text matter more than the count of files in a clip folder.

Coverage quality

Not every mention is equal feedstock. A short wire pickup, a trade deep-dive, and a skeptical FT feature do not carry the same signal, to a human reader or to a system weighting domain and phrasing. Readership-weighted quality and outlet relevance keep the board conversation on which coverage, not only how much. See coverage quality benchmarks.

Message pull-through

Models do not know your message house unless it appears in the sources they read. "We were mentioned" is not the same as "the coverage carried the claims we needed stakeholders to hear." Message pull-through tests whether configured messages appeared inside articles. A generic AI summary of the period will not reliably report what was missing unless that test is explicit.

Narrative framing

Topics are what coverage is about. Narratives are how those topics get characterized. Pricing is a topic; "pricing reflects demand" and "pricing squeezes customers" are two narratives about it. Aggregate volume can rise while the frame most available to compress drifts against you. That is why narrative tracking is a practice rather than a dashboard tile.

Theme-scoped share of voice is the competitive cousin of the same idea: are you present in the conversations strategy funded, or only in the broad category noise? That is the cut in the new share of voice.

What this changes in the weekly practice

You do not need a new department called GEO to act on this. You need a clearer weekly question:

Of the coverage likely to shape how we are described, by people and by systems that compress third-party text, are we on-message, in the right themes, in outlets that matter?

A workable cadence:

1. Name the descriptions you refuse to leave to chance. Three to seven themes or frames tied to OKRs, the same discipline as theme share of voice
2. Define messages as testable claims, not slogans. Pull-through only works if the expected language is specific enough to find or to miss
3. Weight the publication universe. Discovery monitoring can stay wide; the record view for leadership should privilege the outlets that shape category language
4. Read trajectory, not only the latest week. Emerging frames show up in direction across periods
5. Separate upstream evidence from downstream spot-checks. If leadership wants answer-engine appearance tracked, use a purpose-built AI-visibility tool, and bring coverage analysis to explain why the answer looks the way it does

This is still measurement under Barcelona and IEF discipline: objectives first, media outputs clearly separated from audience out-takes, and honest limits on what earned analysis can prove about organizational impact. See how to measure PR ROI in 2026.

What it looks like in a board conversation

Say a head of comms owns category leadership and trust. Overall share of voice is flat. Theme-scoped share of voice shows trust coverage down. Message pull-through on "independent security validation" is weak in the outlets the board reads. Narrative review shows a recurring frame that the category is moving faster than governance.

The board ask stops being "get more coverage so AI notices us." It becomes "fix the feedstock: win trust-theme coverage with the validation message, or the description, human and machine, keeps writing itself without us."

Answer-engine dashboards can later show whether tracked prompts surface it. They cannot invent the coverage quality and framing that never ran.

Where Delve fits, and where it does not

Delve does

• Track earned coverage into projects and structured analysis: topics, sentiment, themes including custom themes, key message mention counts, quotes, publication, and readership
• Support theme-filtered competitive views, so share of voice can sit inside strategic conversations rather than category totals
• Keep message analysis and theme analysis distinguishable, category against claim
• Give teams a continuous upstream view of what was said, where, and whether intended messages appeared

Delve does not

• Rank your brand inside ChatGPT, Gemini, Perplexity, or Google AI Overviews
• Guarantee that on-message coverage will be cited by any given model, since retrieval and ranking are opaque and change
• Treat citation share as proof of reputation or ROI
• Replace judgment about which themes and messages matter

For platform fit on the measurement stack, see best PR measurement tools in 2026. For finding and following coverage first, see best coverage tracking software in 2026.

How to choose your next step