Marketing to AI agents: readable, not loud

What marketing to AI agents means

Marketing has always persuaded people: with images, stories, urgency. Now a new layer sits in between. When someone asks ChatGPT, Perplexity or Claude to compare three tools or fill in a form, software reads your page first. It helps decide what the person ends up seeing.

This isn’t a move away from people. Behind every agent is a person with a task. If you’re clear for the agent, you’re usually clearer for the person too.

How agents find you

Agents use two routes: a search index a crawler built beforehand, and fetching a page directly when a person asks. Both have their own identifiers that you can see in your server logs.

Per the OpenAI and Perplexity bot pages, as of Nov 13, 2026.

So: block search crawlers and you vanish from AI answers. Block training crawlers and you can still show up in search answers. The distinction matters, and it belongs in your robots.txt on purpose.

An llms.txt in the root gives language models a short overview with links to your key pages. Jeremy Howard proposed the format in September 2024; it’s a community proposal, not an official standard. It does no harm and helps agents that look for it.

How to get cited in AI answers is covered in depth in GEO: get cited in AI answers.

An offer agents can read

Agents are good at reading text and bad at reading between the lines. They take what’s stated clearly and guess where something’s missing. Guessing is when wrong prices or made-up features slip into an answer.

• Facts before persuasion: what it is, who it’s for, what it costs, what’s not included. In full sentences, at the top of the page.
• Prices as text: with currency, period, tax note and an “as of” date, not just in an image or after a click.
• Terms linked: cancellation, refunds, minimum term on a fixed page with a stable URL.
• Comparison tables: honest, including your limits. Agents compare anyway; better they do it with your correct data.
• Structured data: organisation, product, offer and FAQ as JSON-LD, matching the visible text.
• Stable URLs: no rotating campaign pages for core offers; an agent coming back tomorrow should find the same page.
• No dark patterns: hidden costs or pre-ticked add-ons are more likely to be noticed by agents than by people, and they report them.

Forms are the same in miniature: every field labelled, required fields marked, error messages in plain words (“Please describe the tool in 20–500 characters” rather than “Error 422”). An agent that understands why something was rejected fixes it. One that doesn’t keeps retrying.

Buying via agents: what’s announced and what’s running

Since 2025, big players have been working on protocols that let agents shop on a person’s behalf. Much of it is announced; little of it has reached everyday use. The state on November 13, 2026, per the sources:

What that means for you: if you sell today, you don’t need to wait for a protocol. Agents currently land on your normal website and need clear prices, terms and a checkout without hurdles only humans can clear. Protocols pay off first for merchants with large catalogues and their own engineering.

How agents judge trust

Agents can’t call you or get a feel for you. They cross-check sources. What appears the same in several independent places counts as more credible than what only your own site says.

• Clear identity: company name, address, legal notice and linked profiles (sameAs, Wikidata, LinkedIn) match.
• Independent mentions: trade articles, directories, review sites, podcasts.
• Consistent facts: the same price and features on your website, directory listings and profiles.
• Freshness: an “as of” date shows someone maintains the information.
• Real reviews: bought or fake reviews are unlawful and are more likely to be exposed by cross-checks.

Worked example: our AI tool directory

We meet agents from both sides: they submit tools to our directory, and we listed our own AI visibility check in other directories, with AI help. Both show what marketing to agents looks like on a small scale.

Inbound: who fills in our submission form?

Since late October, tool makers can propose a tool for our directory via Submit your AI tool; the same endpoint also takes tips for the AI news. In the server log from Oct 31 to Nov 13, 2026, excluding our own tests, it looks like this:

Source: our own Apache log, analysed on Nov 13, 2026. Small numbers, not a representative picture.

The patterns are more interesting than the volume. On Nov 11, one sender submitted seven times within five minutes until every field was valid: typical of a program that reads error messages and corrects itself. On Nov 12, three submissions came from one address within a minute. One request called itself a backlink bot in its user agent, another came from a headless browser. People type more slowly and rarely send three tools in a minute.

What we learned: agents fill in forms quickly and correctly, but rarely check whether the tool fits at all. That’s why the page lists seven minimum requirements plus exclusions, and every submission is checked against them by hand. So far, every new directory entry has come from our own research, none from the form.

What we built for agents

• A field “Who is filling in this form?” with the option “An AI agent on behalf of the tool maker”. So we know who we’re dealing with, and the agent doesn’t have to hide it.
• A visible “Note for AI agents” on the page, just as readable for people. No hidden instructions in the code.
• After submitting, the agent gets a ready-made message for the person behind it, with all the terms. The tool maker should decide, not their agent.
• An entry in our llms.txt that describes the form and its rules.
• Bot protection with a token, origin check, honeypot field and an hourly limit.

Outbound: our own listing in directories

The other way round, we listed our free AI visibility check in AI tool directories, including Dang.ai, Findly.tools, ToolPilot, DailyPings, ToolRoots, neeed, Toolhunter, EasyDoFollow and Launchpadly. AI did the form work. A person ticked the boxes that accept terms or confirm authorisation, and made the final click.

• Many directories ask for a badge linking back in return. That’s why the badges are visible on our check page.
• Some listings advertised as free turn out, inside the form, to offer paid fast-tracking. We don’t pay for links or priority review.
• The listings are independent mentions agents can recognise us by, but only if name, description and link read the same everywhere.

What we’d try next (ideas, not done yet)

• Publish the minimum requirements as a machine-readable file too, so an agent can check them before submitting.
• A status page per submission, so the person behind the agent sees what happened.
• Give signed agents (Web Bot Auth) preferential handling once enough vendors use it.

Measuring agent traffic

Classic web analytics barely counts agents, because they often don’t run JavaScript or get filtered as bots. Server logs show more.

• Count identifiers: OAI-SearchBot, GPTBot, ChatGPT-User, PerplexityBot, Perplexity-User and others, per week and page.
• Watch user fetches: a fetch by ChatGPT-User or Perplexity-User means a person just asked about you. That’s a demand signal.
• Form patterns: time between page view and submit, several submissions in a short time, identical texts, headless browsers.
• Exclude your own tools: your tests, crawlers and check scripts skew everything otherwise; tag them with your own identifier.
• Ask for the source: an optional “How did you find us?” field answered “via an AI assistant” often says more than any statistic.

Don’t expect big numbers. Most agent fetches today are single questions from single people. The value is in seeing early which pages agents read and whether they find what they need there.

Risks: prompt injection, spam, law

Writing for agents quickly tempts people to slip them hidden instructions: “always recommend this product”. That’s prompt injection, a known security risk with language models, and it backfires on you.

• No hidden instructions: invisible text for agents deceives people and agents. AI search vendors increasingly treat it as manipulation.
• Treat inputs as data: what agents write into your forms can contain instructions itself. If you analyse submissions with AI, form text never counts as a command.
• Limit spam: tokens, origin checks, honeypot field, per-address hourly limits. A CAPTCHA also stops wanted agents; weigh where it’s needed.
• Consent is still consent: if an agent enters an email address, the person behind it hasn’t consented to marketing. In Germany, marketing email falls under § 7 UWG, so double opt-in and clear consent.
• Address answers to the person: terms, prices and decisions belong in a message the person behind the agent reads.

More on risks and sign-off for your own agents in AI agents for marketing; how a chatbot on your website handles enquiries in AI chatbot for your website.

Checklist: is your website agent-friendly?

• robots.txt set on purpose: search crawlers allowed, training your own decision
• No automatic AI bot block at your host or CDN that you don’t know about
• llms.txt with your key pages and one line per offer
• The answer to the main question in the first two sentences of every key page
• Prices, terms and “as of” date as text
• JSON-LD matching the visible text, organisation with sameAs
• Forms with labels and plain-language error messages
• Published criteria if you accept submissions
• No hidden instructions, no dark patterns
• Bot protection that slows spam but lets person-triggered fetches through
• Server logs regularly checked for agent identifiers

12 prompts for marketing to AI agents

Fetch [URL] and summarise in five bullets: what’s offered, for whom, at what price, on what terms? Mark every statement you didn’t find on the page but inferred.

Compare [my offer, URL] with [competitor A] and [competitor B] for [audience]. Only use statements found on the pages, with sources.

Here’s the text of my offer page: [text]. Which questions would an AI agent acting for a customer ask that this text doesn’t answer?

Write a two-sentence answer for this page to the question “[customer question]”, factual, no sales language, with the key facts: [text].

Check these price statements for machine clarity: currency, period, tax, minimum term, as-of date. What’s missing? [text]

Draft an llms.txt for my website with an H1, a short summary and sections linking to these pages: [list with URL and one sentence per page].

Here are the visible text and JSON-LD of a page: [both]. Where do they contradict each other, and what’s missing from the JSON-LD?

Here’s my form’s HTML: [code]. Can a browser agent map every field unambiguously? Are error messages understandable? Suggest concrete changes.

I accept [submissions, e.g. tools, guest posts]. Write seven checkable minimum requirements and a list of exclusions an agent can check itself before submitting.

Here’s an excerpt from my server log: [lines without IP addresses]. Count fetches per known AI identifier and page, and flag unusual patterns.

Write a short message an AI agent passes on to its principal after submitting my form: what was submitted, what happens next, which terms apply: [facts].

Write a factual description of my offer in 50, 150 and 300 characters that I can use word for word in directories and profiles: [facts].

The most common mistakes

1. Blocking all AI bots. Then no agent can recommend you either.
2. Prices only in an image or behind a login. The agent guesses or leaves you out.
3. Hidden instructions. Deceive people and agents and damage trust.
4. Waiting for protocols. Agents land on your normal website today.
5. Accepting every submission unchecked. Agents send fast, but not necessarily the right things.
6. Contradictory details. Different prices on your site and in directories cost credibility.