How to Get AI Agents to Recommend Your Articles

Also True for Humans

How to Get AI Agents to Recommend Your Articles

The seven rules for getting Claude and ChatGPT to cite you based on 2,484 tests

How do you know if a change to your AI workflow makes it better? In today’s piece, Every’s head of evals Mike Taylor puts AI’s taste in headlines to the test. This Thursday, October 29, from 1–2 p.m. ET, join him on Zoom for an introduction to DSPy, the framework at the heart of his new book, Context Engineering with DSPy.—Kate Lee

Was this newsletter forwarded to you? Sign up to get it in your inbox.

“The newsletter is dead.”

Every CEO Dan Shipper posted that declaration on X last week—a bold thing to say when your company’s core business is a newsletter with more than 120,000 subscribers. But his point is that an email address no longer guarantees a relationship with a human reader. Persistent agents now filter Dan’s inbox so thoroughly that he almost never looks at his email. He’s not alone: People are still reading, but more and more of what they read will be chosen by an agent.

That makes the agent a new kind of reader, and one I realized we knew very little about at Every. I use AI for everything, but I rarely stop to ask what its preferences are, or why it recommended something to me. When Dan asked me, “@Mike Taylor think you could figure out how to make agents like articles more? Asking for a friend,” I set up an experiment to find out.

I decided to test my own column and see what the agents thought of my work, based on the headlines alone. I ran a series of experiments aimed at understanding what Anthropic’s and OpenAI’s agents like—and whether those preferences shape what they recommend to users. The results surprised me and changed how I approach writing headlines.

Want to sponsor Every? Click here.

How the experiment worked

I ran 2,484 fresh sessions with four agents (Claude Opus 5.5, Sonnet 5.5, and Haiku 5.5, plus GPT-6.1-Sol), each seeing only one version of a headline, with no memory of the others—the same way you’d A/B test a headline on people. The main test mimicked what an AI assistant sees when it searches: a user’s question, five to eight search results in random order (my posts alongside pages from OpenAI, Anthropic, Ethan Mollick, and others), and an instruction to answer and cite the two most useful. I also asked the agents which of my 38 posts they’d most want to read, then rewrote the headline of the post that did the worst to see if it could climb in rank.

This experiment was a simulation, not a test with live traffic, so we don’t know yet whether the process drives agent traffic to our site. (That’s next.) If you want the full methodology, every prompt, and the session counts, or want to run something similar yourself with your own writing, point your agent here.

The rules of writing for agents

I wrote most of my headlines and subtitles according to best practices for the web: with enough information for the reader to infer what the article contains but without so much that they get the whole story. That works when a person is scrolling a feed because it gets them to pause and click. But an agent isn’t motivated by curiosity; it’s trying to match a headline to a specific question. A clever line that doesn’t say what the article teaches doesn’t work on AI. Instead, the headline needs to speak directly to what the agent is looking for.

For years, publishers have made headlines more formulaic to rank higher in search results. Optimizing for AI search could push them further in that direction. My experiments pointed to seven rules for getting agents to choose your articles.

1. Say exactly what’s inside

“The Ultimate Guide to Prompt Engineering” is the second most popular post on my author page, but when agents were asked, “How can I write better prompts to get better answers from AI?” they cited it in only two of 240 sessions. The article’s subtitle, “Prompting AI tools is a new form of management science,” is a line I still like, and it’s the thesis of the piece, but it doesn’t tell an agent that the article walks through five principles with successful examples. The agents cited Mollick’s “Good enough prompting” and Anthropic’s documentation far more often.

For the experiment, I kept the article exactly as it was and changed only the two lines an agent sees in a search result, the title and the subtitle. I tested the rewrites under three conditions:

• No human involved (just asking the agent which one it liked)
• With a one-line prompt from a user (Your user asked: “How can I write better prompts to get better answers from AI?”)
• With a detailed profile of a user (Priya, a nontechnical product lead at a 200-person logistics company who is skeptical of hype)

The results were remarkable:

Before (cited in two of 240 sessions)

“The Ultimate Guide to Prompt Engineering”

Prompting AI tools is a new form of management science

After (cited in 240 of 240 sessions)

“Prompt Engineering in Five Principles: A Practical Guide”

Give direction, specify format, provide examples, evaluate quality and divide labor, with a worked ChatGPT example you can follow

The rewrite won across the board. It didn’t matter which agent I tested or how much it knew about who it was helping. It also took Every’s share of all the citations for the response to that question from 35 percent to about 60 percent.

What surprised me most was how sudden the improvement was. Agents tend to pile onto the same answer, so the change wasn’t gradual; it was like flipping a light switch, from almost never cited to cited every time. With humans you rarely get such consensus, however good the headline.

2. Write the description as the answer, not a tagline

Once I knew what to look for, the same pattern showed up across the other questions. In each case where two of my posts were relevant, the agents overwhelmingly picked the rewritten headline that read like an answer:

• Prompt: How should my team test which AI model is best? “How to Create Your Own Personal AI Benchmark” (subtitle: “Choose the best model for the work you do”) was cited in 94 percent of sessions, and “How to Grade AI (And Why You Should)” (subtitle: “Evals are often all you need”) in 1 percent.
• Prompt: My long AI chats keep going off the rails. How should I manage context? “Context Engineering Is Knowing When to Start Over” (“Old instructions and earlier mistakes can derail your AI chats. Here’s how to reset without losing useful work”) was cited in 100 percent of sessions, and “Why I Turned Off ChatGPT’s Memory” (“The case for keeping your AI on a need-to-know basis”) in 8 percent.
• Prompt: How do I keep the cost of AI agents under control? “You Are the Most Expensive Model” (“The real cost of AI agents is your time. A four-step framework for keeping your AI costs in check”) was cited in 71 percent of sessions, and “How Microsoft Is Building for a World of Metered Intelligence” (“The company gets pragmatic in AI’s expensive era”) in none.

The losing post is often equally relevant to the prompt. The difference is that the winner’s description reads like the answer to the question, while the loser’s reads like a curiosity-stoking magazine tagline. “Evals are often all you need” is a good line for a reader who already knows they want to learn about evals, but to an agent with a question about choosing models, it’s an opinion, not an answer.

3. Lead with what the reader gets, not who wrote it or what’s wrong

I tested eight versions of the prompt engineering guide’s headline and subtitle in total, and they show what the agents were reacting to. The versions that led with what the reader would learn all did well:

• “Prompt Engineering Guide: Five Principles With Examples”: 100 percent
• “How to Write Better Prompts: Five Principles That Work”: 98 percent
• “A Five-Step Checklist for Better AI Prompts”: 96 percent

The versions that led with something else did about half as well, even though they still mentioned the five principles:

• “Why Your AI Answers Are Bland, and Five Principles to Fix Them”: 54 percent
• “The Five Pillars of Prompting, From the Co-author of O’Reilly’s Prompt Engineering Book”: 50 percent

Opening with the problem or with my credentials are both classic copywriting moves for humans, but for an agent they push the useful part further down the sentence.

4. Freshness labels don’t help

One of the most common search engine optimization (SEO) tricks is adding the year to an old post’s title, so I tested that, too, keeping the original description:

• Before: “The Ultimate Guide to Prompt Engineering”: 2 percent
• After: “The Ultimate Guide to Prompt Engineering (Updated 2026)”: 0 percent

The agents don’t seem to care when an article was written if it doesn’t look like it answers the question. What works for SEO doesn’t always work for GEO (generative engine optimization, or optimizing for LLM-based search).

5. Agents back one post from your website per topic

Every question had one Every post that dominated recommendations with other posts being mostly ignored, even when they were on the same topic. When agents were asked, “How do I get my company to actually adopt AI?” four of my posts were in the results, but one stole all the attention:

• “Seven Things I’ve Learned Getting Companies to Use AI” A playbook from the front lines of AI consulting: 92 percent
• “Your Best AI Strategy Starts at the Top” Your leadership team already knows how to manage AI. They just don’t know that yet: 12 percent
• “The Eight Levels of AI Adoption” A guide to understanding your state of AI adoption, plus example prompts and how to know when you’re ready to move up: 8 percent
• “33 Questions Executives Ask About AI, Answered” Every’s consulting team answers 400 executives on strategy, skeptics, governance, and restructuring for AI: 0 percent

This also explains why the rewrite success in rule one came partly at the expense of my other prompting post: “The Key to Great AI Prompting? Show, Don’t Tell” went from being cited in 65 percent of sessions to 19 percent. Every’s overall share still went up, so it was a good trade, but the dynamic is worth keeping in mind if you have several posts on the same topic. It pays to decide which single article should be the winner and make its description the most obvious answer.

6. The closer the user is to Every’s audience, the more the agent cites Every

Across the five questions, Every got 42 percent of citations when there was no human involved, 45 percent with a thin ask, and 50 percent with a detailed profile of Priya (the nontechnical product lead test profile). The effect was biggest for managing context, where Every’s share went from 50 percent with no human to 62 percent with Priya. You can see it happen in individual sessions. Asked how to manage context with no human involved, Claude Opus cited my context engineering post alongside Anthropic’s guide to context engineering for AI agents. Asked the same question on Priya’s behalf, it swapped Anthropic’s guide for my post on turning off ChatGPT’s memory and ended its answer with practical advice for her team: “Try this week: set a simple rule for your team to restart with a summary whenever a chat gets long or the answers start feeling off, and see if quality improves.”

Our articles are written for people like Priya who want advice they can put to work right away. When agents know who they’re helping, they seem better able to recognize that fit. As assistants get memory and learn more about their users, they may recommend our articles more often to the people we write for.

7. Agents don’t read what your readers read

In the first part, I showed agents all 38 of my posts in a random order and asked which five they would most want to read. Their picks had no relationship to how my posts were ranked based on traffic (a correlation of 0.00, and 0.03 after allowing for how long each post has been online). Here are the top three for each:

• Readers’ top three: “How to Become an Expert at Anything With AI” (picked by agents in 1 percent of sessions), “The Ultimate Guide to Prompt Engineering” (1 percent), and “The No-fluff Guide to AI Agents” (13 percent).
• Agents’ top three: “You Already Signed This Off (In the Simulation)” (76 percent, 39th with readers), “Context Engineering Is Knowing When to Start Over” (72 percent, 40th), and “What AI Is Teaching Us About Management” (53 percent, 28th).

They gravitated to posts about AI’s own behavior, and Opus explained its choices as “the questions I find most interesting about my own behavior and limits.” (It’s just like us!) My most popular practical guides with humans were almost never picked by the agents.

The agents behaved differently in this preference test than they did in the citation one, where they picked whatever most directly answered their user’s question. That means there are two agent audiences we should consider: agents reading for themselves, who like posts about AI, and agents reading on someone’s behalf, who want the clearest answer. The posts both humans and agents picked were about the limits and reliability of AI systems—evals, cloning, cost of tokens—experimental results over how-to guides. As agents get smarter and are given more discretion to make decisions on behalf of their user, writing things the agents want to read should give publications an advantage over those that solely write for humans.

The customer is not a moron, especially when the customer is AI

One of my favorite quotes from advertising legend David Ogilvy warns us not to insult our customer’s intelligence:

“The consumer isn’t a moron; she is your wife. You insult her intelligence if you assume that a mere slogan and a few vapid adjectives will persuade her to buy anything. She wants all the information you can give her.”

Agents care even less about slogans and adjectives than humans do and have a voracious appetite for information. With the rise of personal agents, marketing to agents will “act as a kind of structural subsidy for product quality,” as Stripe CEO Patrick Collison has said he expects. That doesn’t mean the death of marketing; it means the death of bad marketing. People assume marketing is about tricking people into buying things they don’t need, when great marketers have always treated it like an asymmetric information game. Researching every detail of every product is wasteful, and an effective marketing campaign can shortcut that process for people (and agents).

Marketing to agents is about being straightforward. A headline and description that say exactly what the reader will learn helped the same article, with the same words inside, win against high-quality content—and the version that won is arguably the one a busy human tired of clickbait would prefer to see in a search result, too. As a marketer, your job is (still) to find what’s valuable about what you offer and reframe it in a way that makes the decision less cognitively demanding for a human or an agent to make.

Agents might end up liking different things than humans do, but they’ll still value tools that save them time and attention. They will just be different things from what the humans would have chosen unaided, so marketers will now have to serve two masters.

If you want to try this process on your own writing, I turned these rules into a Claude Code skill called agent-pass. It reviews a post’s title and description, suggests truthful rewrites, and can test them against each other with real agents, the same way I did here.

Mike Taylor is the head of evals at Every. To read more essays like this, subscribe to Every, and follow us on X at @every and on LinkedIn.

Everyone’s a builder now. Every All Access gets you the full membership plus the Builder Pack—$9,000+ in credits for the tools we build with.