I was researching a product recently and asked an AI assistant to compare a few options. Instead of an answer, I got confusion. Different retailers were presenting different details about the same product, so the assistant could not confidently tell me which one to buy.
In that moment, the brand lost. Its own facts were sitting on other people’s pages, in versions that disagreed, and none of them was authoritative. If that brand had published one definitive version of its own product information, the assistant would have had a way through the conflict, and something to actually tell me.
That moment is worth taking apart.
Your website sells by implication
Humans buy with emotion. Put the same wine into two different bottles and people will judge it by how classy or exclusive the bottle looks. Put similar products from different brands side by side and you may feel more drawn to one than the other.
That pull is what a website is built to create, the same way a physical retail space is. The design is trying to help a person answer two questions: am I the sort of person who buys this product, and will people approve of me if I buy it?
Look at the same product on two different websites and one will land better. Is it luxurious and stylish? Is it for people who research deeply before they buy? Is it good value, decent things at fair prices? That positioning comes through in the imagery, the colours, the fonts, the language. The whole package, landing at once.
Even the anti-brand is a package. Some of the newer skincare brands sell on transparency: a short ingredient list, one ingredient that works rather than fillers that sound good and do nothing. That appeals to the buyer who thinks: I don’t fall for the big brand nonsense. Different feeling, same mechanism.
And you never say any of it out loud. You communicate by implication.
The agent doing your buyer’s research can’t see the package
What changed is that buyers now turn to AI for part of their research. Someone asks an assistant to suggest products that meet their requirements, or to compare the options they have shortlisted. The assistant reads whatever public surfaces it can reach and reports back. That is happening now.
I am not claiming agents are making purchases. That is still speculative, and the argument does not need it. Research is enough, because research is where the shortlist gets formed.
An agent doing that research can’t see the package. It does not infer premium positioning from a serif or trustworthiness from white space. It reads text, and it needs that text to be explicit.
This argument is already running in developer circles. Kurt Mackey, the founder of Fly.io, wrote recently that a carefully curated developer experience with opinionated defaults may now be a liability, because “agents work best when things are explicit.” He was talking about APIs and documentation. I think the same logic lands harder on marketing, because marketing has spent a decade getting better at exactly the wrong technique for this reader: compression, distillation, don’t make me think, one confident claim instead of five qualified ones. Every one of those is a human-audience instinct, and every one of them removes what the agent came for.
Your polish may be the problem.
Keep the crafted page, and add an explicit one for agents
The instinct will be to optimise one page for both audiences. I think that serves neither. Strip the voice and imagery out to make a page machine-legible and you trade a working asset for a speculative one. Leave it as it is and the agent keeps arriving at a page built to be felt rather than read.
So ship two artifacts. Keep the human page exactly as crafted as it is. Alongside it, build a second, deliberately unglamorous surface for the agent: full spec, real numbers, pricing and limits, materials or ingredients, what it is for, which integrations exist, comparisons that admit where you are not the right choice. Explicit and structured, so it can be read directly rather than inferred.
Then attach the proof, because this is the half most teams will skip. An agent weighing you against a competitor on verifiable facts is doing something no human buyer ever did at that stage of research. An unsupported claim does not just fail to persuade it. The agent may discount the claim entirely. Every claim on the agent-facing surface should have a number, a customer, a third-party validation or documentation sitting next to it. If the proof does not exist internally, that is a finding worth escalating, not something to word around.
This is not the “optimise for AI search” advice that is everywhere right now. That is a findability argument, and it assumes your content is fine and just needs to be more discoverable. The claim here is different. Agents can find your content. It does not contain what they need, because it was built to work as a package for someone who experiences it all at once.
Agents present your stale pages as current, and reward fresh ones
One more mechanism, and this is the one that turns a content task into a risk. An agent reading your pricing page has no way to know the page is stale unless the page tells it. It will summarise last year’s packaging as current, with your name attached. When people found your pages themselves, an out-of-date page was a hygiene issue. When an agent reads it for them, it becomes a misrepresentation of your company that you did not authorise and cannot see.
It cuts the other way too. Agents take recency into account, precisely because of this problem. An agent is likely to prefer a page that says when it was last reviewed over one nobody has touched in a year. So the ongoing refresh does two jobs: it protects you from misrepresenting yourself, and it shows the agent that you are the one giving it correct, current information.
Either way, the agent-facing surface cannot be a one-off project. Six months stale, it is worse than not existing.
Four ways the two-artifact approach fails
- The wall of text. A team reads “agents want explicit information” and publishes an unstructured dump of spec text. That helps nobody. It reads as low quality to a human who stumbles on it, and it gives the agent no more structure than the original page had. Explicit is not the same as undifferentiated.
- The merge. Someone decides two artifacts is duplication and folds them into one compromise page, which is the failure the whole recommendation exists to avoid.
- The orphan. Nobody owns the agent-facing surface, so it quietly decays into the misrepresentation problem above.
- The overclaim. This is a bet on where agent-led buying goes. How buyers’ agents will actually read marketing pages is genuinely unsettled, and a team that builds this may see nothing measurable for a couple of quarters. I think the bet is good, because the cost is low and the downside of being wrong is small. But it is a bet, and anyone telling you otherwise is selling something.
What the two-artifact approach needs in place
- You know which of your public surfaces agents actually read for your category. You find that out by running the research yourself, not by assuming.
- Someone owns the machine-legible layer as a named responsibility, with the last-reviewed date visible on the page. It may need refreshing more often than the human page does, because showing the agent it has been recently reviewed is itself part of the signal. The product release cadence is the minimum, not the target.
- The team has explicit permission to maintain two artifacts, so the merge instinct can be resisted when it arrives.
- Your top claims have evidence that can be attached.
Product marketing is the natural owner. One named person runs a monthly audit: ask three AI agents to go and read the pages, then answer the questions your buyers actually ask. It takes under an hour and it is the cheapest diagnostic available.
Where to start
-
Run the audit this week. Ask three AI agents to visit your site and answer the questions your buyers ask: should I use you or your top competitor, what do you cost, do you integrate with X, what are your limits, who is this not for. Tell them to read the pages rather than answer from memory, and run the comparison against a competitor’s live site too. Write down every wrong or missing fact. That list is a content backlog, and the first run always finds something.
-
Build one explicit twin. Take your most-read product page and produce a second version with no compression: full spec, pricing, limits, integrations, a source behind each claim. Make it crawlable. Do not touch the original.
-
Attach proof to your top five claims. The number, the customer, the validation, the documentation. Any claim where the proof does not exist is worth escalating.
-
Name the owner. One person accountable for what agents say about the company, with a monthly hour booked. If nobody currently holds this, that sentence alone is worth bringing to your next leadership meeting.
The research moment I opened with is the cheapest version of all of this. One question to an assistant, and a brand that had left its own facts to chance. Whether yours reads the same way is an hour of work to find out, and I would spend the hour.