Key takeaways
- Press share of voice counts coverage you earned. AI share of voice counts how often models name or cite you when buyers ask category questions. A lead in one does not predict a lead in the other.
- News coverage is a short-lived source for ChatGPT. In Promptwatch Data, half of cited news pages were gone after 36 days, while landing pages and docs stayed cited far longer.
- AI share of voice varies a lot by engine. Pooling ChatGPT, Perplexity and AI Overviews into one number hides the gaps.
- Read both numbers side by side per topic, then act on the gap: a press lead with weak AI visibility is a content problem, and the reverse is an earned media problem.
- You can't buy AI share of voice with media spend. Content depth and third-party validation move it.
Why PR teams now need two share of voice numbers
PR teams have used share of voice for decades. You count your brand's mentions in media coverage, divide by total mentions for you and your competitors, and you get a percentage. It is a clean way to benchmark presence after a launch, a campaign, or a news event.
Then buyers started asking ChatGPT, Perplexity and Google AI Overviews who to shortlist. The press number still matters, but it no longer describes the whole conversation. A brand can dominate trade coverage and barely appear in AI answers. The opposite happens too, and it is the more uncomfortable case for a comms team: a competitor with thin press gets recommended constantly because its product pages and docs are easy for a model to cite.
So the job is not to replace one metric with the other. It is to read both, understand why they diverge, and know which lever to pull.
What each metric actually measures
Press share of voice
The inputs are articles, broadcasts and social posts. The formula:
Press SOV = your mentions / total category mentions (you + competitors) x 100
It rewards volume, reach and placement in outlets people read. Spend, pitching and news events all push it up. Many teams weight it by outlet tier or by sentiment, which is sensible.
AI share of voice
The inputs are AI-generated answers to a defined set of prompts. The common formula:
AI SOV = your brand mentions in answers / total brand mentions across tracked prompts x 100
A few definitional choices change the number a lot:
- Per mention or per answer. Counting every mention rewards a brand that is named three times in one response. Counting once per response is steadier. Promptwatch counts one mention per response per brand. GripIQ also argues for per-answer counting.
- Mentions or citations. A mention is the model naming you. A citation is the model pointing to a source as evidence. HG Insights makes the point that a high mention count with a low citation rate means the model knows you but doesn't trust you enough to source you.
- Pooled or per engine. 5WPR gives a simple example: if a brand is named 180 times in 1,800 pooled mentions, its AI share of voice is 10%, but that single number can hide very different results in each model.
That last point matters for PR readers. A 10% pooled figure could be 25% in one engine and 2% in another, and those two situations need different responses.

The structural difference: you can't buy it
HG Insights puts it bluntly: AI share of voice can't be purchased the way media spend can. A model cites what it judges to be structured, verified and trustworthy, so the route to a higher score goes through content depth and third-party validation.
I think this is the part PR leaders resist most, because it breaks the usual budget logic. More pitching raises press share of voice almost mechanically. It raises AI share of voice only if the resulting coverage ends up on pages models actually retrieve and cite, and if your own pages back it up.
What the citation data says about press coverage
This is where the two metrics really split. Promptwatch Data looked at what ChatGPT cites and how long those pages stay cited.
Press is a short burst
In Promptwatch Data's ChatGPT content half-life study, every page ChatGPT cited on July 6, 2026 was tracked until October 4, 2026. News and press coverage decayed fastest: almost a quarter of news pages were never cited again after day 0, and half were gone after 36 days. At day 60, 41% were still cited.
Compare that with owned and evergreen formats from the same study:
| Content type | Still cited at 30 days | Still cited at 60 days |
|---|---|---|
| Landing pages | 78% | 69% |
| Docs | 76% | 67% |
| Comparisons | 72% | 63% |
| Listicles | n/a | 60% |
| Product pages, reviews, how-tos | n/a | about 55% (half-life 74 to 77 days) |
| News | about 50% at day 36 | 41% |
A big launch week can lift your press share of voice and your AI share of voice together, and then the AI side fades quickly if the lift came mostly from news pages. Promptwatch's own advice from that data is to use news and press coverage for short bursts of visibility.
What ChatGPT cites overall
Promptwatch Data's content type report (Sep 28 to Oct 4, 2026) shows product pages at 22.4% of classified ChatGPT citations and landing pages at 16.8%. Listicles were 13.6%, how-tos 8.7%, comparisons 5.9% and news 5.7%. Roughly 39% of citations go to pages describing what a company sells.
For branded prompts, where a buyer asks about you by name, reviews were 5.6% of citations, more than three times their overall share. Press releases were 2.1%, which is above their share across all prompts but still a small lever. If your PR plan leans on the wire, that number is worth a look.
Engines differ, so pooled numbers mislead
Press is not the only thing that varies by engine. Promptwatch's social media citations data for Oct 11 to Nov 9, 2026 found social platforms were 13.16% of Google AI Overviews citations, 1.94% of Perplexity's and 0.64% of ChatGPT's. AI Overviews cites social media about 20x as often as ChatGPT. YouTube alone was 6.71% of AI Overviews citations.
So a campaign built around video and community conversation may move AI Overviews and do almost nothing in ChatGPT. Read per engine, or you will misattribute results.
How to read the two numbers together
Put both on the same topic list, over the same period, against the same competitor set. Then look at four patterns.
| Pattern | What it usually means | What to do |
|---|---|---|
| High press SOV, low AI SOV | Coverage is earned but models don't retrieve or trust your own pages, or the coverage sits in short-lived news | Strengthen landing pages, docs and comparisons. Check crawler access. Get coverage on listicles and review sites, not only news |
| Low press SOV, high AI SOV | Your owned content is strong, competitors are louder in media | Convert that content into pitchable data and commentary. Press coverage may add third-party validation |
| High on both | Healthy, but check per-engine gaps | Watch for decay after campaigns end |
| Low on both | Awareness and content problem | Start with the topics where buyers ask category questions |
Before acting, check mentions against citations. Mentions without citations suggest a trust or content problem. Check sentiment as well, because a mention that frames you as expensive or limited is not the win a raw count implies.
Building a combined measurement routine
1. Fix the prompt and topic set
Use the same themes in both systems. Presspage's approach of breaking AI share of voice down by theme is useful here, because it lets you see which subjects your brand is associated with rather than only counting mentions. Write prompts the way your buyers do, including comparison and branded variants.
2. Choose a counting rule and keep it
Pick per-answer or per-mention and never change it mid-year. Consistency matters more than which you pick.
3. Track per engine
Report ChatGPT, Perplexity, Gemini, AI Overviews and others separately, plus the pooled figure. The pooled number is for the board slide. The split is for decisions.
4. Trace the sources
For each topic where you trail, look at which pages the model cites. If competitors win through third-party listicles, that is a placement target for PR. If they win through documentation, that is a content job.
5. Report decay, not just level
For each campaign, record AI share of voice at launch, at 30 days and at 60 days. The half-life data suggests news-driven lifts will fall faster than lifts from evergreen pages.
Tools for tracking AI share of voice
The press side is covered by Meltwater, Cision and similar media monitoring suites, which I haven't compared here. For the AI side, there are many options.
Promptwatch tracks share of voice, visibility score, position and sentiment from the real interfaces of ChatGPT, Gemini, Perplexity, AI Overviews and others, and traces which pages and third-party sources the models cite. Its metrics overview and Share of Voice definition explain how it counts. Beyond tracking, it includes crawler logs, content gap analysis and a Content Agent for fixing gaps, plans start at $95 per month and are confirmed at promptwatch.com/pricing.

Other tools worth a look, depending on your setup:
Profound is an enterprise AI search analytics platform. See Promptwatch vs Profound for a head-to-head.
Semrush's toolkit suits teams already in that suite.
Peec AI is a lighter prompt-tracking option.

Otterly.AI is an affordable monitoring-focused choice.
| Tool | Strength | Consider if |
|---|---|---|
| Promptwatch | Tracking plus citation tracing, crawler logs and content fixes | You want to act on the gap, not only report it |
| Profound | Enterprise analytics | Budget is large and monitoring is the priority |
| Semrush AI Visibility Toolkit | Fits an existing SEO workflow | You already pay for Semrush |
| Peec AI | Simple prompt tracking | You need basic share of voice numbers |
| Otterly.AI | Low cost | You are testing the metric |
Shadow, mentioned in the research, positions itself as a communications operating system that combines media, social, AI visibility and SEO reporting. That integrated view appeals to comms teams, though I haven't verified its feature depth. For a wider set of options, the GEO software directory at bestgeosoftware.com lists many more.

Pitfalls to avoid
- Don't compare a pooled AI number to a press number and call it a verdict. Different engines, different source mixes.
- Don't treat a launch spike as a trend. The decay data says news lifts fade.
- Don't count mentions you wouldn't want. Check sentiment and the claims being made.
- Don't assume a press lead protects you. Models retrieve pages, not reputations.
- Don't skip the technical side. If AI crawlers can't read your pages, no amount of coverage fixes that. Promptwatch's AI crawler logs show whether bots read a page and where they fail.
A simple monthly scorecard
For each of your top five topics, record:
- Press share of voice and the top three outlets
- AI share of voice per engine and pooled
- Mention versus citation rate
- Sentiment score
- Which source types the models cite when you lose
Then assign each gap an owner. Press gaps go to media relations, citation gaps go to content, and access problems go to the web team. A shared dashboard helps, since the usual failure is that PR and SEO each hold half the picture.
If you'd rather hand parts of this off, agencies such as 1001 SEO Media run GEO and digital PR together, which is relevant when the answer to a gap is both a content fix and a placement.
The bottom line
Press share of voice tells you how loud you were in published coverage. AI share of voice tells you whether models name and cite you when buyers ask. They diverge because models favor structured, durable pages and treat news as a quick hit. Measure both by topic and by engine, watch decay after campaigns, and use the gap between them to decide whether your next dollar goes to a pitch or to a page.

