Something strange started showing up in server logs and lead sources over the last couple of years: sessions that behave like extremely efficient researchers. They hit the pricing page, the FAQ, the comparison content, and the contact details in seconds, no meandering, no images loaded. Some of these are AI agents doing research on behalf of a human who asked their assistant to "find me three options for X."
I do not know exactly what share of any given site's traffic is agentic. Nobody honestly does. But the direction is unmistakable, and I have started designing web presences with a second audience in mind: software that reads on behalf of buyers. Here is how I think about it.
The Shift: From Persuading Readers to Informing Retrievers
Human-first marketing relies on tools that do not work on machines: atmosphere, social proof carousels, emotional pacing, the slow build of a landing page narrative. An agent parsing your site is doing something closer to structured extraction: What is this? What does it cost? Who is it for? What are the constraints? Can I verify these claims?
If your site makes those questions hard to answer, the agent does not get persuaded anyway. It moves to a competitor whose facts were extractable, and its human never even sees your brand. The failure is silent, which is what makes it dangerous.
This is a continuation of what I wrote about in preparing websites for AI crawlers, but agents raise the bar: crawlers index you for later, while agents evaluate you in real time with a purchase decision attached.
What I Actually Change on Sites Now
1. Put the answers where the evasions used to be
The classic B2B move of hiding pricing behind "book a call" is increasingly expensive. When an agent is assembling a shortlist with budget constraints, "pricing not disclosed" is not intrigue, it is disqualification. I now push clients toward publishing at least pricing structures, ranges, and starting points, along with plain statements of who the offer is and is not for. The pages that sales teams resisted writing for years turn out to be exactly what the agent economy rewards.
2. One page per fact, one fact per claim
Agents (and the retrieval systems behind assistants) work best with pages that have a clear, single job: a real FAQ, a specifications page, a shipping and returns page, a "how our process works" page. Sprawling pages that mix poetry with facts extract badly. I keep the brand storytelling, but I make sure every commercially important fact also exists somewhere in unambiguous, liftable prose, the same discipline that drives LLM citations.
3. Structured data as an API you did not have to build
Product, Offer, Service, LocalBusiness, and FAQ schema are no longer just rich-result bait. They are the closest thing most businesses have to a machine-readable catalog. When an agent can pull your price, availability, and attributes from JSON-LD instead of guessing from prose, you have reduced its cost of considering you. Reduced friction wins shortlists.
4. Keep facts consistent everywhere, because agents cross-check
An agent comparing your site against your marketplace listings, your Google profile, and third-party reviews will notice contradictions instantly, at a thoroughness no human shopper matches. Price mismatches, outdated hours, conflicting claims: each one is a trust deduction. Consistency maintenance used to be brand hygiene. It is becoming conversion optimization.
The Trust Layer: Verifiable Beats Impressive
Here is my strongest opinion on this shift: in a web flooded with AI-generated claims, provenance becomes the differentiator. Agents are being built, sensibly, to prefer sources that can be verified: reviews on platforms with fraud controls, claims corroborated by third parties, data with a traceable origin.
That means the marketing assets that age best are the boring, verifiable ones: real review volume accumulated honestly, editorial coverage, published data from your own operations, consistent entity information across the web. I go deeper on this thesis in first-party truth and data provenance, but the short version is that the era of asserting your way to credibility is closing. Machines check.
What Does Not Change
I want to keep this honest, because "optimize for AI agents" can slide into hype fast.
Humans still make the final decision in almost every purchase that matters, and they still choose between the shortlisted options using entirely human criteria: taste, trust, brand affinity, how the sales conversation felt. The agent narrows; the human picks. So brand building, positioning, and the emotional work of marketing lose none of their value. What changes is that you now have to survive a machine screening round to reach the human round.
Think of it like hiring: a resume parser rejects you before any person reads your cover letter. The cover letter still matters. But only if you pass the parser.
A Practical Starting Checklist
If you want to act on this next week rather than think about it next year:
1. Ask three different AI assistants to research your category and recommend providers. Note whether you appear, and what facts they get wrong.
2. Read your own pricing, FAQ, and service pages as an extraction task. Every question a shortlisting agent would ask should have a plain-text answer.
3. Validate your structured data and expand it to cover offers and FAQs, not just articles.
4. Audit fact consistency across your site, profiles, and listings.
5. Check that your robots and firewall rules are not blocking the agent user-agents you actually want (many sites block them all by default and never revisit the decision).
The buyers are starting to bring software to the research phase. The sites that treat that software as a legitimate audience, rather than an intruder to block or a fad to ignore, are going to be on shortlists their competitors never knew existed.