How to use AI for product marketing

Ask a product marketer how they use AI and you will usually hear about writing: first drafts of launch copy, battlecard updates, summaries of a pile of call recordings. That is useful, and it is also the smaller change. The bigger one is that many B2B websites now have an AI agent on them that answers buyers in its own words, all day, to everyone who asks. Your positioning used to be something a rep delivered on a call. Now software delivers it, and most PMMs have never read what it says.

I call this the most-spoken pitch: the version of your story buyers hear most often is no longer the one in the positioning doc or the one your best rep gives. It is whatever the agent says when a stranger types "how are you different from X?" at 11pm. The PMM should own those words the way they own the homepage.

What does AI do for product marketing?

AI for product marketing splits into three jobs, and they need very different amounts of a PMM's attention.

The first two jobs have a human between the AI and the buyer. The third does not. That is why it deserves the most care and usually gets the least. Most of the product marketing tool stack still sits in the first two columns.

Why is the website agent the most-spoken pitch?

Because most of the buying committee never hears your rep. Storylane's Modern Buying Trends 2026 report, built on 8,211 buyer conversations across 19 companies and 637 deals, found that nine in ten of the buying committee never join a sales call. It also found that 51.4% of conversations with website agents start outside the buyer's working hours. The pitch your enablement team rehearses reaches a minority of the people who decide, and the battlecards it trains on rarely reach the buyer at all.

A static website says exactly what you wrote. An agent answers in its own sentences, rewriting your positioning once per question. When those answers go off-message, the usual cause is not a weak model. The agent says what its sources say, and most of its sources were written by people who were not thinking about positioning.

Where does a website agent get its words?

Mostly not from the pages product marketing writes. The Storylane Buyability Index asks B2B websites the questions a buyer would ask and records which page supplied each full answer. The homepage supplied just 1.8% of full answers. On comparison questions ("how is this different from..."), a dedicated comparison page supplied only 19.7% of the full answers. Blog posts, help centres and docs supplied far more, which I wrote about as answer distance. One caveat: the questions were generated for each company rather than logged from real buyers, and the sites are not a random sample of B2B.

So when an agent answers a competitor question, it is more likely paraphrasing an old blog post than the comparison page you signed off. When it answers a product question, it is often working from a help article written for existing customers. The words came from whoever wrote last, not from whoever owns the message. The full breakdown is in the report on what B2B websites answer, half-answer and dodge.

What do buyers actually ask a website agent?

Less of what marketing spends its time on, and more of what marketing leaves to sales. Internal Storylane data (RepX website conversations) hand-coded 2,225 buyer questions across three RepX deployments (the coded subset in Storylane's Modern Buying Trends 2026 report). Questions could carry more than one label.

Pricing and capability come to 42%. At larger scale, Storylane's Modern Buying Trends 2026 report found buyers are 32% more likely to raise pricing and capability with a website agent than with a human rep, which is one more reason to decide whether to put pricing on your website. Customer proof and the ROI story, two things PMMs pour weeks into, come to 2% each. Buyers rarely ask for proof, so it has to arrive inside the answer to the question they did ask.

Competitor questions are rare but they matter most. In the same data, the 62 conversations where a buyer named a rival ran 10.3 buyer messages on average against 4.5 for the other 748, and qualified at 53.2% against 36.0%. That is a small sample, so treat the figures as indicative. The direction held at all four vendors with enough of these conversations to measure. More detail is in five learnings from 1,332 RepX conversations.

So your agent's answer to "how are you different from X?" is read by fewer people than your pricing answer, but they are among the most qualified visitors on your site.

How should a PMM train a website agent?

Treat it like onboarding your sharpest new rep, except this rep will repeat whatever you teach them to every visitor. Five decisions belong to the PMM.

1. Choose the sources. Decide which decks, call scripts, recordings, docs and pages count as current, and remove the old launch deck that still describes last year's packaging. The agent cannot tell an outdated slide from a current one. You can.
2. Write the positioning in plain English. Who the product is for, what it replaces, and the one or two things you want every answer to carry. Write it the way you would brief a rep, not the way it reads on the homepage.
3. Write the competitor answers yourself. For each rival that comes up, a fair answer: where you fit better and when the buyer might reasonably pick the other one. Buyers comparing vendors can tell when an answer dodges. Borrow from how your reps already do objection handling.
4. Set guardrails. Rules the agent must never break: how it talks about pricing, what it will and won't say about competitors, compliance claims it can't make.
5. Match the message to the page. A visitor on the pricing page and one in the docs are at different moments. The opening prompt, the call to action and the qualification questions should differ by page (more in page-level context for conversational AI).

None of this is new work for a PMM. The difference is that an agent uses all of it, on every conversation, instead of leaving it in a folder.

How can a PMM use agent conversations as research?

Every question a buyer types is an unprompted piece of voice of customer research: their words and their order of concerns. A website survey asks what you want to know. An agent records what they wanted to know. It complements win-loss analysis: win-loss tells you why deals ended, the agent log tells you what buyers wondered on the way in.

Here is a weekly rhythm I'd suggest (a suggestion, not something we measured).

Half-answered questions are the best signal in the pile: they tell you exactly which claim is vague. Those fixes belong on the page as much as in the agent, so send this list to whoever owns website messaging research, usually the same PMM.

Shouldn't demand gen or RevOps own the agent?

That is the strongest objection, and half of it is right. The agent books meetings and routes leads, so its routing rules, qualification thresholds and CRM handoff belong to RevOps and demand gen. I wrote about this split in who owns the website AI agent.

But routing is not the message. Nobody would let RevOps write the homepage because the homepage has a demo button on it. The words the agent says are positioning, and positioning is the PMM's job. So RevOps owns where the conversation goes, and the PMM owns what it says.

Where RepX fits

I work at Storylane and we make RepX, a website agent, so read this section with that in mind. RepX is built for teams where product marketing wants to own what the agent says. It learns from your website, decks, sales call recordings or transcripts, Storylane demos and FAQs. FAQs give a fixed, consistent answer for pricing, compliance wording and marketing-approved messaging, which is where a PMM's exact words go. Qualification runs on your ICP written in plain English. Objection handling, including competitor comparisons, is trained on how your best reps respond, and guardrails on pricing, competitors and compliance are taught once and enforced everywhere.

Matching the message to the page is what Playbooks do. For each page you set the starter questions, intro message, CTA buttons, qualification questions and when to ask for an email. Each interactive demo gets a description and a plain-language display instruction that controls when RepX shows it, so the buyer sees the real product mid-conversation (the reason the interactive demo and AI agent combo works). You can test the agent before launch and check changed answers with evaluations.

For the research loop, RepX analytics show engagements, qualified conversations and booked meetings, filtered by intent, with drop-offs and high-interest topics. Storylane MCP lets you ask Claude, ChatGPT or Copilot what buyers asked and which accounts to follow up, with a weekly report. A Zapier trigger fires when a conversation completes, for example to brief a rep before outreach. Qualified leads and conversation summaries flow to HubSpot, Salesforce and Slack. RepX starts at $2,000 a month, and the price depends on website traffic.

Hospitable shows what this looks like when a PMM runs it. Their product marketing manager trained the agent on sales call transcripts from every kind of rep, the help centre and the demo library, and piped every conversation into a shared Slack channel so the team could read answers and correct them. RepX booked 36 meetings and sourced 7 closed-won deals in its first month at Hospitable, and he says it handles about 80% of the questions they get.

What to do this week

Open your agent's or chat tool's conversation log and read the last twenty answers it gave to a competitor or pricing question. For each, ask: would I let a new rep say this? If not, trace the words back to their page and fix them there. That one hour will tell you whether your most-spoken pitch is the one you wrote.

If you want to see how RepX handles your own messaging and competitor questions, book a demo.