Playbook · Reviewed October 2026
GEO for comms teams: who owns it and how to start
A playbook for PR and communications teams on generative engine optimization: why comms belongs in the work, how to split it with SEO, marketing, legal, IR and product, a 90-day starting plan, and what to measure.
Comms owns the story; SEO owns the site
Generative engine optimization (GEO) is the work of improving how often and how accurately AI assistants describe your organization. For most organizations, comms should own the benchmark questions, the accuracy of answers and the earned-media program, while the SEO or web team owns the technical side. Legal, IR and product supply and approve facts.
The reason is practical. Much of what assistants cite is press coverage and other third-party material, and comms already manages those relationships and the facts the organization states publicly.
AI answers draw on earned and owned media
Assistants that search the web cite pages, and many of those pages are news stories, reference sites and the organization’s own site. Muck Rack’s Generative Pulse report (May 2026 edition, ChatGPT, Claude and Gemini, more than 25 million links across 17 industries) attributed 84% of citations to earned media. Its definition of earned media is broader than journalism, which on its own accounted for 27% of cited links.
Our own citation-sources study looks at which domains five assistants cite for corporate and reputational questions, measured separately from Google AI Overviews. The two pieces of work use different samples and definitions, so read each with its own scope.
Press coverage and AI visibility are still different measurements. A strong quarter of coverage does not tell you what an assistant says, and the only way to know is to ask it. See our glossary entries on earned media and owned media.
A starting RACI
R is responsible for doing the work, A is accountable for the outcome, C is consulted and I is informed. Adjust it to your organization; the point is that every row has one A.
| Activity | Comms / PR | SEO / web | Marketing | Legal | IR / public affairs | Product |
|---|---|---|---|---|---|---|
| Benchmark question set | A/R | C | C | I | C | C |
| Monitoring answers | A/R | C | I | I | C | I |
| Owned fact pages | A | R | C | C | C | C |
| Structured data and llms.txt | C | A/R | I | I | I | C |
| Earned media program | A/R | I | C | I | C | C |
| Correcting errors | A/R | R | I | C | C | C |
| Reporting to leadership | A/R | C | C | I | C | I |
A 90-day starting plan
Weeks 1–2
Baseline
- Name one accountable owner and agree the RACI with each team.
- Write 20 to 40 benchmark questions and freeze the wording.
- Run them across ChatGPT, Claude, Gemini and Perplexity, more than once each, and record Google AI Overviews separately.
- Log every error and every cited domain.
- Inventory the owned pages that state core facts.
Weeks 3–6
Fix and publish
- Correct or redirect owned pages that state outdated facts.
- Add or update Organization structured data.
- Request corrections from publishers and reference sites, following their rules.
- Note the questions where assistants cite no source you would choose, and plan coverage for those topics.
- Send in-product feedback on high-severity errors.
Weeks 7–12
Measure and report
- Re-run the same question set and compare the two snapshots.
- Report visibility, cited domains and accuracy, with dates and sample sizes.
- Set a monthly cadence and an escalation route for high-severity errors.
- Decide whether the manual process is enough or whether you need tooling.
For the correction work in weeks 3 to 6, our playbook on correcting what AI assistants say about your company includes an evidence-log template and a guide to tracing errors to their sources.
What to measure
Use measures you can define and repeat. Each figure needs its date, the assistants sampled and the number of answers.
| Measure | What it tells you | Watch out for |
|---|---|---|
| Visibility | The share of sampled answers that mention your organization. | Not share of voice. Varies between runs, so report the sample size and dates. |
| Cited domains | Which websites the answers link to or name as sources. | Assistants differ in how many sources they cite, so compare like with like. |
| Share of citations | A domain’s share of all citations in the sampled answers. | Pooled shares lean toward the assistants that cite the most sources. |
| Accuracy of descriptions | Whether answers state your facts correctly, checked against your own fact sheet. | Needs human review. Log errors by type and severity, as in the correction playbook. |
PeakMetrics is a narrative intelligence platform that shows communications, public affairs and risk teams how stories about their organization spread across news, social media and AI answers, and flags when those narratives are coordinated or manipulated. Its AI Perceptions product runs benchmark questions daily across ChatGPT, Claude, Gemini, Perplexity and Grok and reports the domains each answer cites. The AI Perceptions free check publishes dated visibility and citation snapshots by sector.
Common mistakes
Treating one answer as the truth
Live answers vary between runs. Decide on a sample, then judge the pattern.
Changing the questions every month
New wording means a new measurement. Retire and add questions deliberately, and keep the rest fixed.
Pooling Google AI Overviews with the assistants
AI Overviews is a search surface with different sources. Report it separately.
Leaving GEO to the web team alone
Technical fixes help crawlers read your site. They do not correct a wrong news story or brief a spokesperson.
Counting mentions without checking accuracy
A high visibility figure can include answers that describe you wrongly.
Editing Wikipedia directly
Wikipedia’s conflict-of-interest guideline strongly discourages it. Use the talk page and disclose your affiliation.
When an answer is invented rather than sourced, see AI hallucination for why it happens and what you can and cannot change.
Frequently asked questions
Who should own GEO, comms or SEO?
How is GEO different from SEO?
How many benchmark questions does a comms team need?
Does press coverage affect what AI assistants say?
Is AI visibility the same as share of voice?
Product names are trademarks of their respective owners. Third-party figures are as published by their source and reviewed in October 2026; please tell us if anything here needs correcting.
Measure press coverage and AI answers side by side.
See how ChatGPT, Claude, Gemini, Perplexity and Grok describe your organization, next to the news and social narratives behind them.