How to market when agents do the research.

Marketing that works through AI agents, in two directions. One is marketing to the research and shopping agents that gather evidence, compare vendors and assemble shortlists for buyers. The other is running your own marketing with agents that work between human checkpoints. The first now defines the term, because that is where buyers moved.

The shift is measurable. In G2’s March 2026 survey of 1,076 B2B software buyers, 51% said they now start research with an AI chatbot more often than with Google, up from 29% in April 2025; 69% chose a different vendor than they had planned because of chatbot guidance, and a third said an AI chatbot led them to choose a vendor they hadn’t considered. By Cloudflare’s count, fewer than half of all HTML page requests now come from a human, though not every machine is an agent working for one.

What changed is who gets on the shortlist, not when it forms. Shortlists always formed before a seller was called; agents changed who is eligible to be on them. And visibility isn’t one thing: a page can be retrieved, cited, recommended, chosen and paid for, and each is a separate result with separate causes. Map the decision each question serves before you measure anything.

Much of what’s sold under the label deserves suspicion. Special agent files, voice rewrites and single visibility scores have the weakest evidence behind them. The unglamorous work holds up better: pages an agent can read, facts it can check, and one consistent description of the company.

From the pages they can fetch, for the constraints the buyer gave them. An agent turns a request like “SOC 2 automation for a 50-person company, under $20k” into several searches, retrieves many pages and cites a few: in AirOps’ March 2026 analysis, ChatGPT ran two or more follow-up searches on 89.6% of queries and left 85% of the 548,534 pages it retrieved uncited. What gets you considered is specific and checkable: pricing logic, named integrations, honest comparisons and outcome proof.

The list still forms early; agents change who is on it. In 6sense’s 2025 buyer research, the winning vendor was already on the day-one shortlist 95% of the time, and 94% of buyers ranked that list before speaking to a seller. By the time the shortlist is filled, buyers already know 75% of the vendors on it; with about five evaluated, that leaves roughly one open seat for a vendor they haven’t dealt with. An agent’s research is one of the few ways a new name takes it.

Access isn’t attention. In a 2026 simulation of agent-mediated shopping, growing from 6 to 24 platforms raised the chance a product was in the candidate pool by about 8 points but cut its shortlist inclusion by 13.6 points and its purchases by 13.0. Being available, shortlisted, first and bought are separate states. And an empty category is a lost one: if you’ve never published in it, the answer is assembled from whoever did.

Usually the first HTML response, and only what’s in it. In Vercel and MERJ’s crawler study, none of the major AI crawlers rendered JavaScript; Gemini, on Googlebot’s infrastructure, and AppleBot were the exceptions. ChatGPT’s crawler also spent 34.82% of its fetches on pages that returned a 404, against 8.22% for Googlebot. Fetch your key pages as raw HTML: if the prices, integrations or proof arrive only after scripts run, most crawlers never see them.

Agentic browsers are a different case: they load pages the way a browser does, and vendors don’t document what their agents take from them. Design for the first response anyway, keep one name and description everywhere, and never serve agents a different page from the one people see, which Google treats as cloaking.

Access is a policy, set per purpose. OpenAI runs four user agents: OAI-SearchBot for ChatGPT search, GPTBot for training, ChatGPT-User for visits a user asks for, and OAI-AdsBot for pages submitted as ChatGPT ads. Opting out of OAI-SearchBot keeps you out of ChatGPT search answers; blocking GPTBot doesn’t. Decide search, training and user-requested visits separately, as most sites already do: fewer than 1% of Cloudflare’s sites block search bots, while 17% block training in some way.

Agent standards are barely adopted. Of the 200,000 most-visited domains Cloudflare scanned, 78% had a robots.txt, 4% declared AI usage preferences and 3.9% served Markdown to agents that asked for it. Google says its AI features need no special files. llms.txt is an optional hour: nothing measures a root-level file, and the work it can’t replace comes first.

Evidence it can check, and less than you’d hope. Models can be moved by claims nobody verified: when researchers put fictitious products in a model’s prompt, unsupported social-proof lines like “best-selling” raised how often it recommended a product, while scarcity and exclusivity framing lowered it. In another 2026 study, fabricated clinical-evidence claims broke a well-known brand’s hold on recommendations. Read that as a threat model, not a tactic.

So treat every popularity or proof claim as a governed fact, with a source, a denominator, a date, an owner, an approval and an expiry. Buyers notice when it’s missing: in G2’s survey, 64% said they encounter inaccurate AI chatbot recommendations often or very often.

And a model’s confidence proves nothing. In a 2026 audit of four LLM recommenders across three catalogs, each model’s stated confidence stayed in a narrow band while its catalog accuracy ranged from about 39% to 99%, and filtering on that confidence changed the hallucination rate by at most 1.65 points. Paid placement changes the price of prominence, not what makes a page citable: ChatGPT now carries ads, and OpenAI checks their landing pages with its own crawler. Keep paid and organic in separate columns.

The work you can check, with approval wherever a mistake is expensive. Delegate by consequence: sort each action into run, review, approve or prohibit, and put the human checkpoint at publishing, spending, contacting a customer and making a claim, not at drafting. How independently an agent works and what it’s allowed to change are separate decisions.

Oversight is visibility plus the power to step in, not a click on every step. In Anthropic’s study of 998,481 sampled API tool calls, 80% came from agents with at least one safeguard and 73% appeared to have a human in the loop, figures Anthropic calls upper bounds, and 0.8% of actions looked irreversible. Experienced users of its coding agent auto-approved more, from about 20% of sessions to over 40% by their 750th, and also interrupted more, from 5% to about 9% of turns.

The gains are real at the task level. Across 453 professionals, ChatGPT cut writing time by 40% and raised quality by 18%; in BCG’s field experiment, consultants were faster and better inside the model’s range and 19% less likely to be right outside it. Nothing yet supports whole-program claims, so automate observation and drafting, and keep the claims and the publish button with a named person. What a CMO should keep follows the same rule: positioning, competitor claims, pricing language and public statements.

The bottleneck moves from drafting to review. When a model drafts in minutes, what caps a team is how much work each person can be accountable for. Give every recurring artifact a named claim owner and a separate approver, budget reviewer hours the way you once budgeted writers, and cap the work in flight. Don’t spin up a separate AI team: the people who own the claims own the agents.

Adoption is outrunning governance. The CMO Survey found AI now performs 17.2% of marketing efforts, and 81% of leaders in Microsoft’s 2025 Work Trend Index expect agents to be moderately or extensively integrated into their AI strategy within 12 to 18 months. Yet in the IAB’s State of Data 2025, only 33% of agencies, brands and publishers used or planned mandatory human oversight of AI outputs, and 21% an AI governance board or point person.

The scarce skills are judgment: choosing the question, substantiating the claim, declining to publish, and reading results against predictions you wrote down first. Vibe marketing works when it runs as a loop with a review stage and an approval gate, not as a way around them.

In separate columns: what engines say, what agents fetch, who arrives and what they buy. Answer observations need the same questions repeated across phrasings and engines, with their spread. Fetches belong in server logs and prove eligibility, not demand. Referrals are partial: GA4’s AI Assistant channel counts arrivals from sources like ChatGPT, Gemini and Copilot but leaves Google’s AI Overviews and AI Mode in Organic Search. Revenue needs a declared attribution rule, and every claim a label: observed, attributed, inferred or causal.

A fall in clicks isn’t a fall in demand. With an AI Overview on the page, Ahrefs measured a 58% lower average click-through rate for the top-ranking result, and in Pew’s study users clicked a traditional result on 8% of visits with an AI summary against 15% without. The click was a proxy for being considered: measure presence on a fixed question set, ask how buyers heard about you at the form, and name what you can’t observe.

And activity isn’t buyers. In Profound’s sample of 7.5 million ChatGPT conversations, the share classified as commercial rose from 13.9% to 19.2% in a year; the widely quoted “more than doubled” is Profound’s extrapolation from estimated weekly users, not a count. Only 16% of brands track AI visibility systematically, by the IAB’s count, so a clean baseline is still an advantage.