Key takeaways
- You can't control the model. You can control the evidence it finds: your own pages, reviews, third-party coverage and a few social platforms.
- Track narrative at the claim level ("affordable for small teams", "weak support"), not just mentions and a visibility score.
- Branded prompts lean on product pages and landing pages (about 23% each), so your own site is the first place to fix a wrong story.
- Which engine you're dealing with changes the plan: social barely moves ChatGPT, but YouTube dominates AI Overviews and Reddit plus LinkedIn dominate Perplexity.
- Run it as a loop: capture, score, diagnose, fix, re-test. A 30-day cycle is enough to see movement.
Why PR teams now own a new surface
A story about your company used to reach people as a headline. Increasingly it reaches them as a summary. According to the Reuters Institute's 2026 Digital News Report, 10% of people (16% of under-35s) use AI chatbots for news each week, up from 7% a year earlier. And the same summaries answer the questions buyers, candidates and investors ask when they research you.
That summary is stitched together from several outlets, your website, review sites and whatever else the engine has read. You didn't write it. Nobody approved it. Yet it's probably the most-read description of your brand.
So the PR job has grown. You still pitch, respond and monitor coverage, but you also need to know what the combined picture says, which pages produced it, and how to change it.

What "brand narrative" means in AI answers
A mention is a yes/no. A narrative is the set of claims an AI repeats about you. "Affordable for small teams." "Its mobile app lacks features." "Good support, slow onboarding." Those claims shape whether someone shortlists you, and they can be wrong, stale or borrowed from a competitor's framing.
It helps to separate four things that get lumped together:
- Visibility: are you mentioned at all, and how prominently?
- Sentiment: is the tone positive, neutral or negative?
- Accuracy: are the specific claims true?
- Source: which page taught the model that claim?
Most dashboards stop at the first two. PR teams need the last two, because accuracy and source are what you can actually act on. I'd argue a brand that is mentioned in every answer with the wrong pricing is in worse shape than one that's occasionally absent.
To be blunt about the limits: you cannot force ChatGPT, Gemini, Claude, Perplexity, Copilot or AI Overviews to repeat a fixed message. You can audit what they say, find the sources behind weak claims, fix contradictions and re-test until the answer improves.
The five-step narrative control loop
- Capture what the major engines say across real buyer prompts.
- Score each answer for accuracy, sentiment, citations and recommendation quality.
- Diagnose the source, omission or ambiguity behind each weak answer.
- Fix the evidence: your pages, reviews, third-party coverage.
- Re-test the same prompts until the answer improves, or the risk is documented.
Step 5 is the one teams skip. Without a fixed prompt set you can't tell whether your fix worked or the model just rolled the dice differently.
Step 1: build a prompt set that mirrors how people ask
Start with 30 to 50 prompts, split into three types:
- Branded: "Is [Brand] good for small teams?", "What does [Brand] cost?", "[Brand] reviews"
- Competitor comparison: "[Brand] vs [Competitor]", "alternatives to [Competitor]"
- Category: "best [category] tools for [persona]"
Write them the way real buyers, journalists, candidates and investors would. A CFO and an intern ask differently. Run the same set on a schedule, across every engine your audience uses.
Step 2: score every answer
A simple rubric works fine in a spreadsheet to begin with:
| Dimension | What you check | Score |
|---|---|---|
| Accuracy | Are pricing, features, audience and facts correct? | 0 to 2 |
| Sentiment | Tone toward the brand | Negative / neutral / positive |
| Citations | Is your own domain among the sources? | Yes / no |
| Recommendation | Are you recommended, listed, or ignored? | 0 to 2 |
| Message match | Does the answer reflect your key messages? | 0 to 2 |
Tag each wrong or weak claim with a theme (pricing, features, support, reputation). Patterns show up fast. Usually three or four claims account for most of the damage.
Step 3: diagnose where the claim came from
Every weak answer has one of three causes:
- A source says the wrong thing (an old pricing page, a bad review, an outdated article).
- No source says the right thing (an omission, so the model guesses).
- Sources conflict (the model hedges or picks the loudest one).
Look at the citations behind each answer. That's the PR equivalent of tracing a rumor to its origin.
Which sources actually shape the narrative
This is where the data should drive your priorities. Promptwatch Data analysed ChatGPT citations by content type for Sep 28 to Oct 4, 2026. For branded prompts, the ones that name your company or product:
- Product pages made up 23.3% of citations and landing pages 22.8%. Together that's almost half.
- Reviews were 5.6%, more than 3x their share across all prompts.
- FAQ pages were 4.8%, social posts 3.0% and press releases 2.1%.
So when someone asks ChatGPT about you by name, your own pages are the main text it reads, and reviews punch well above their weight. Press releases still show up, a little more than average, but they're a small slice.
For commercial prompts, product pages (26.0%), landing pages (20.6%), listicles (15.0%) and comparisons (7.9%) make up almost 70% of citations. Translation: third-party listicles and comparison posts are the PR targets that matter for "best X" questions.
One more thing for media relations. News decays fast. In Promptwatch's content half-life report, news pages lost half their citations in about five weeks, while landing pages and docs held most of theirs for months. A launch story gives a short bump. Your evergreen pages carry the narrative.
Social platforms: engine by engine
Social is where one-size-fits-all advice fails. In Promptwatch's November 2026 social citations report (Oct 11 to Nov 9, 2026, a 30-day window compared with a 7-day October report, so treat changes as directional):
| Engine | Social share of citations | Top platforms |
|---|---|---|
| Google AI Overviews | 13.16% | YouTube 6.71%, Facebook 2.34%, Reddit 2.10% |
| Perplexity | 1.94% | Reddit 0.82%, LinkedIn 0.66% |
| ChatGPT | 0.64% | Reddit 0.40%, LinkedIn 0.20% |
AI Overviews cites social about 20x as often as ChatGPT, and YouTube alone out-cites the other six platforms combined there. For Perplexity, Reddit and LinkedIn are about three quarters of social citations. X is negligible everywhere (0.02% in AI Overviews).
What that means in practice:
- If ChatGPT matters most to your buyers, don't spend your narrative budget on social posts. Fix your site, reviews and third-party comparisons.
- If you care about AI Overviews, a YouTube presence (demos, explainers, customer stories) is a real lever.
- If Perplexity is where your audience researches, watch Reddit threads and LinkedIn posts that mention you.
Reddit's role also moves around. Promptwatch recorded reddit.com's share of ChatGPT citations falling from about 4% to 0.5% on August 14, 2026, and it has since recovered to 0.40% of ChatGPT's social-platform mix in the November window. Don't build a strategy on one platform's behavior this quarter.
Fixing the evidence
Once you've traced weak claims to sources, the fixes are fairly unglamorous.
Fix your own pages first
Because product and landing pages dominate branded citations, make sure the homepage, product pages and docs answer the questions people actually ask: pricing, features, who it's for. State facts in plain sentences. If your pricing page is a "contact us" wall, the model will fill the gap from somewhere else.
Add a plain-language FAQ page. FAQs are 4.8% of branded citations, small but cheap to get right, and they let you word claims exactly the way you want them read.
Also check the technical basics. If AI crawlers can't read a page, nothing on it can shape an answer. Promptwatch's free Agentic Readiness Check shows which AI crawlers can reach your site, and the robots.txt generator for AI helps if you've been blocking bots by accident.
Fix the third-party layer
- Reviews: collect them on the trusted sites in your category. They're over 3x more prominent in branded citations than in general ones.
- Listicles and comparisons: pitch the authors of the "best X" articles AI already cites for your category. This is traditional media relations pointed at a new target list.
- Corrections: when a specific article carries an outdated claim, ask for an update. A polite email to an editor still beats everything else.
Rethink your media list
Media lists are normally ranked by reach, authority and relationships. AI search adds another factor: which articles engines draw on for your category. A trade publication with modest traffic can show up in answer after answer. Rank outlets by how often they're cited, not only by readership.
Tools for tracking brand narrative
You can run the first cycle by hand. Asking nine engines twenty questions every week and comparing the answers doesn't scale, though, so most teams move to a platform quickly.
Promptwatch is the one I'd start with for this job. Its Brand Narrative feature pulls individual claims out of every tracked response, groups repeated ones, shows which models say each and files them under themes like pricing, features, support and reputation, with a 0 to 100 score per theme you can compare against competitors. Fact Check (Business plan and up) lets you mark a claim accurate or inaccurate, then reads your crawled pages to find the ones that state it or its opposite. That's step 3 of the loop, done for you. It also collects from the real interfaces of ChatGPT, AI Overviews, Gemini, Perplexity and others, which can differ from API output. Check promptwatch.com/pricing for current plans; Brand Narrative is on every plan, and there's a 7-day free trial.

Profound tracks how the main AI engines describe a brand, including sentiment and citations, and offers agents built on its prompt data. It's a strong option for enterprise comms teams.
Profound

Otterly.AI and Peec AI are lighter prompt trackers. They answer "was I mentioned?" reasonably well, but they won't tell you why or help you fix it.
Otterly.AI

For the older half of the PR stack, you'll still want media monitoring. Meltwater and CisionOne cover traditional coverage and earned media, which feeds the evidence AI reads.
| Need | Tool type | Examples |
|---|---|---|
| Claim-level narrative, fact checking, fixes | AI visibility platform | Promptwatch, Profound |
| Basic mention tracking | Prompt tracker | Otterly.AI, Peec AI |
| Traditional coverage and earned media | Media monitoring | Meltwater, CisionOne |
| Social listening | Social intelligence | Brandwatch, Talkwalker |
If you're comparing options, the GEO software directory lists more of them.
A 30-day plan
Week 1: baseline
Build the prompt set, run it across engines, score answers with the rubric and list the top five wrong or weak claims. Put the numbers in a one-page report. You'll need that baseline for step 5.
Week 2: diagnose
Trace each weak claim to its sources. Sort them into wrong source, missing source and conflicting sources. Rank by how often the claim shows up and how much it could cost you (a wrong price hurts more than a bland adjective).
Week 3: fix
Update your own pages first, since that's fastest. Then start the slower work: review collection, outreach to listicle authors, correction requests, FAQ publishing.
Week 4: re-test and report
Rerun the identical prompts. Compare against the baseline. Don't expect every claim to flip, because engines refresh at different speeds and news-derived claims decay on their own. Document what changed, what didn't, and the residual risk. Then repeat monthly.
Mistakes I see most often
- Treating a mention as a win. Check what is said and whether it's accurate.
- Averaging across engines. ChatGPT, AI Overviews and Perplexity pull from very different source mixes, so one blended score hides the problem.
- Over-investing in press releases. They're a small slice of branded citations (2.1%).
- Changing the prompt set every month. You lose the ability to compare.
- Skipping the owner. Assign one person to the narrative report, or it becomes everybody's side project.
If you'd rather not run this in-house, an agency that handles both SEO and AI search work, such as 1001 SEO Media, can set up the monitoring and the content fixes together.
The takeaway
Narrative control in AI answers is an evidence system. Pick the prompts that matter, score what comes back, trace each weak claim to its source, fix that source, and re-test with the same prompts. For branded questions, the biggest lever is boring: accurate product pages, landing pages and FAQs on your own site, backed by reviews and the third-party listicles AI already trusts.


