How to Watch the Narrative in a Category

Watching a narrative means tracking what AI says about a market rather than who it names. Write down the claims the market is supposed to stand for, add the awareness-stage questions that pull those claims out, track them daily, then follow the citations back to the publishers the story is coming from. Five steps, and the last two are where a comms team finds work it can actually do.

A market ranking answers "who." A narrative answers "how," and the how is usually what a communications team is paid to move.

Here's the version of this that shows up in real life. A trade body or a category leader notices that AI has started describing their market in terms nobody in it would choose. Maybe the cost objection gets repeated as fact. Maybe an older technology gets treated as the default and the newer one as the risky option. Maybe a safety concern that was settled five years ago keeps coming back. Nobody wrote that story, and it's being repeated to every buyer who asks a general question about the category.

The story came from somewhere. Models don't invent a market's framing; they assemble it from sources, and a handful of sources do most of the work. Finding those sources is the point of this workflow.

This guide works with any AI visibility setup; the callouts show how each step runs in Spyglasses.

Step	What you're doing	What tells you it's working
1. Write the claims	Turning the story into trackable statements	Three or more claims, one idea each
2. Cover awareness	Adding the questions asked before the shortlist	All three buying stages represented
3. Track daily	Putting the prompts on a schedule	A chart per claim instead of a hunch
4. Point goals at publishers	Attaching outlets to a tracked project	Scored outlets with a goal each
5. Read the sources	Following the claims back to their origin	A short list of domains to work on

Add the claims you expect AI to repeat

A market's story is made of a few claims. Write them down, because a claim nobody wrote down can't be tracked, and "the narrative feels off" isn't something you can put in front of a board.

Start with what the market is supposed to stand for. If your industry association has a positioning line, it's usually three claims wearing one sentence; split it. Keep each claim to a single idea in plain language, the way a buyer would say it rather than the way a press release would.

Then write down the claims you'd rather AI dropped. This half gets skipped and it's the more useful one. An outdated cost figure, a safety worry that was resolved, a comparison to a technology the market moved past. You want to know the day one of these starts surfacing, and you can only be told about claims that exist as records.

Three claims is a working floor. Eight is about as many as anyone reviews regularly. If your list is longer than that, some of them are the same claim in different clothes.

Cover the awareness stage

Claims live at the awareness stage, and most prompt sets are thin there.

That's understandable. Awareness prompts feel unsatisfying the first time you read the answers, because the buyer doesn't know the market well enough to ask for a comparison. They ask about the problem, not the product, so the answer comes back as an explanation with few brand names in it or none at all.

That explanation is exactly what you're here for. It's the market's story in the model's own words, and it's the thing being read by every buyer who hasn't formed a shortlist yet.

So add the questions buyers ask before they know the options. "How do companies handle this at scale?" "What are the approaches to this problem?" "What goes wrong with the usual way of doing this?" Keep the consideration and decision prompts too, since the ranking still matters and the claims turn up there in a harder-edged form. Tag the awareness set so you can read it on its own; mixing the three stages in one view hides the pattern you're looking for.

Track the claims every day

A narrative moves slowly. That's the whole reason to read it daily.

One reading tells you almost nothing, because AI answers vary from session to session and any single answer might be the unusual one. Thirty readings tell you which claims are holding, which are fading, and which one started climbing three weeks ago. The variation stops being a problem and turns into a baseline you can measure against.

Put the category prompts on a daily schedule under a project named for the market, and set a start date you can point at. From then on, read two things.

Pull-through per claim, which is how often the claim you wrote down actually shows up in the answers. Expect this to be low at first and to move in single digits. A claim that goes from appearing in 8% of awareness answers to 15% over two months is a real shift, even though neither number sounds impressive.

Then the claims you didn't write. Skim the answers themselves every few weeks, not just the numbers. New framing shows up as language before it shows up as a metric, and a claim that isn't on your list can't be counted yet.

Point goals at the publishers that shape the category

A few outlets set the terms of any market. Once you can see which ones the answers lean on, outreach stops being a list of everybody who covers the industry.

Shortlist the outlets the market's answers cite. Then score them before you commit anyone's hours, because two things decide whether an outlet is worth working. Whether AI can read it at all, since plenty of well-known publications block AI crawlers outright, and whether it already shows up for questions in this market. A famous masthead that blocks crawlers won't move an answer no matter who you know there. Building a pitch list AI can actually see covers the scoring in depth.

Then attach the ones you're working as goals on the tracked project. A goal ties the outreach to the measurement, so when a claim's pull-through moves you have a specific piece of work sitting next to it rather than a guess. This is the same mechanism that connects earned media to AI visibility for a brand; here the subject is the market instead.

Read which sources AI trusts

The last step is the one that turns all of this into a plan.

Every tracked answer carries its sources, so the domains cited most often for this market are sitting there in a list. Read it and compare it against the outlets you already work with. Three patterns come up almost every time.

Outlets you know and already have relationships with, which is the easy case and usually the first place to take a correction.

Sources you didn't expect, which is the interesting case. Reference works, standards bodies, well-moderated forums, and trade publications far outside the tier list. These are frequently easier to correct than a magazine feature, because they're maintained rather than published once.

And sources carrying the claim you want changed, which is where to start. Sort the work by which source is feeding the framing you're trying to move rather than by the size of the outlet. A mid-sized trade site that AI reads constantly beats a national title it can't read at all.

That's the whole payoff of watching a narrative rather than a ranking. You don't end up with a number to report; you end up with four domains, a claim each, and a reason to call them.

What a quarter of this gets you

Narrative work is slow and it compounds, so judge it by the quarter rather than the week. Three months in, a category being watched properly has four things.

1. A claim list with history. Three to eight claims, each with a pull-through line long enough to have a normal range. The range is what makes an unusual month readable as one.
2. An awareness set that works. Enough questions at the top of the journey that the market's story shows up in the data rather than in somebody's inbox screenshot.
3. A source list, ordered. The domains AI reads for this market, scored, with the ones carrying the claims you care about at the top.
4. One correction you can point at. A claim that moved after specific outreach, with the dates attached. One cause and effect beats a quarter of charts.

If you haven't set the market up yet, mapping a category in AI answers is the five-step version of the setup, and this workflow picks up where it ends. If the same market is also a new-business list, turning a category map into prospects uses the ranking rather than the claims. And if the story has turned against the category quickly rather than drifting, crisis communications when AI is repeating the story is the faster-moving version of this work.

Part of the Spyglasses workflow series. This guide is one of twelve workflows for PR and communications teams. The rest are collected in the AI visibility guides for PR and communications teams.