Every: how to get AI agents to recommend your articles
As AI assistants increasingly sit between a search query and the articles that answer it, a practical question has emerged for writers and editors: what makes an AI agent choose your piece over the alternatives? Mike Taylor, head of evals at Every, published findings in October 2026 from an experiment designed to answer that question.
The methodology
Taylor ran 2,484 independent sessions across four AI agents: Claude Opus 5.5, Claude Sonnet 5.5, Claude Haiku 5.5, and GPT-6.1-Sol. Each session received a headline variant in isolation, with no memory of other tests — structured as an A/B test. The primary experiment gave each agent a user question and five to eight search results, then asked it to cite the two most useful. A secondary experiment asked agents which of 38 posts they would most want to read themselves. This was a simulation, not a live-traffic test, so results describe agent behavior rather than confirmed traffic effects.
The seven rules
Say exactly what’s inside. Describe the specific content, not the thesis. One rewrite — making the title and subtitle explicit about what the piece actually contains — took a guide from cited in 2 out of 240 sessions to cited in every session.
Write the description as the answer, not a tagline. Descriptions that directly answered the question the article addresses outperformed descriptions that framed the topic or introduced the author.
Lead with what the reader gets. Opening with the problem being solved, or with the author’s credentials, consistently weakened citation rates. Opening with the outcome performed better.
Freshness labels do not help. Adding “(Updated 2026)” to a title drove citation share to zero in the test. Agents appear to treat explicit freshness signals as distraction rather than quality markers.
Agents back one post per topic. When multiple articles covered the same question, one absorbed most citations and the others received very little. The implication for publishers is to pick the single strongest piece on a topic and optimize it rather than spreading effort across multiple overlapping articles.
Audience proximity raises citation share. When the agent was given a detailed user profile that matched the article’s assumed reader, citation share rose from 42% to 50%. Writing for a specific audience, and stating that audience clearly, helps agents match content to questions.
Agents do not read what your readers read. The correlation between article traffic and agent citation was close to zero. Agents reading for their own interest preferred posts about AI behavior and limitations. Agents answering a user’s question preferred the most direct answer regardless of traffic rank.
What this means for writers and editors
The main practical finding is that clarity in headlines and metadata carries more weight with AI agents than it does with many human readers. Taylor argues this is not a compromise: writing that answers the question directly serves both audiences. The study does not address on-page content structure, only titles and descriptions, so it is a starting point rather than a complete optimization guide.
Writers and editorial teams working on long-form content, newsletters, or knowledge bases will find the citation-share methodology worth adapting for their own testing.