How to Get AI Agents to Recommend Your Articles
Why this is in the vault
Mike Taylor (Every's head of evals) ran 2,484 live sessions across Claude Opus 5.5, Sonnet 5.5, Haiku 5.5, and GPT-6.1-Sol to test which headline/description patterns get an AI agent to cite an article over competitors — a direct, data-backed GEO (generative engine optimization) playbook, not a vendor's theoretical pitch.
The core argument
Headlines written for human curiosity (a teaser that withholds the payoff) actively lose against headlines written as literal answers, and the effect is binary rather than gradual — "like flipping a light switch." Taylor's own post "The Ultimate Guide to Prompt Engineering" went from being cited in 2 of 240 sessions to 240 of 240 after rewriting only the title/subtitle (unchanged body) to state the five principles directly: "Prompt Engineering in Five Principles: A Practical Guide."
Seven rules fell out of the experiment:
- Say exactly what's inside — vague/clever titles lose to literal ones.
- Write the description as the answer, not a tagline — "Evals are often all you need" (opinion) loses to "Choose the best model for the work you do" (answer).
- Lead with what the reader gets, not the author's credentials or the problem being solved — credential-first and problem-first framings scored roughly half as well as benefit-first framings on the same facts.
- Freshness labels don't help — adding "(Updated 2026)" to a title did nothing; SEO tricks don't transfer to GEO.
- Agents back one post per topic — when multiple of Taylor's own posts covered the same question, the agents consolidated nearly all citations onto a single "winner," even cannibalizing a previously-winning post of his own. Implication: pick one canonical piece per topic rather than letting several compete.
- The closer the user profile resembles Every's actual audience, the more agents cite Every — citation share rose from 42% (no user context) to 50% (detailed persona), suggesting agents increasingly match content to audience fit as they gain memory/context about the user.
- Agents reading "for themselves" (no user to serve) pick different posts than either human readers or agents serving a user — they gravitated to meta content about AI's own behavior/limits, with near-zero correlation (0.00–0.03) to human readership rankings.
No third-party sponsor in this issue. Standard Every house self-promo in the footer (Every All Access bundle: Sparkle/Cora/Spiral/Monologue) — self-promotion, not a disclosed sponsorship, and doesn't bias the experiment's findings. Zero deep-fetches taken: the newsletter's only cited third parties (Mollick's "Good enough prompting," Anthropic's context-engineering docs, a Patrick Collison tweet) are used as comparison data points inside Taylor's own argument, not a curated list of separate recommended reads — nothing met the "specific hook + plausible RDCO relevance" bar for a follow-link beyond what's already summarized here.
Mapping against Ray Data Co
This directly operationalizes RDCO's stated interest in being found/cited by AI systems as part of building credibility for phData sales (project_credibility_for_phdata_sales): Sanity Check pieces and the phData-facing credibility content should be titled and subtitled as literal answers to the question a reader (or an agent on a reader's behalf) would ask — not as curiosity-hooks — if the goal is agent citation/recommendation traffic rather than just human scroll-stop. Rule 5 (one canonical post per topic) is a direct warning against Sanity Check running multiple takes on the same angle, which the vault's own feedback_no_derivative_sanity_check_pieces memory already independently arrived at from a different direction (dilution of original reframing) — this piece supplies the citation-share mechanism for why that's true, not just the editorial-taste reason. Rule 6 (agent citation share rises with audience-profile specificity) also reinforces positioning Sanity Check narrowly rather than broadly, since a sharper reader profile appears to improve agent-citation odds, not just human relevance. The author also shipped the underlying method as a reusable Claude Code skill (agent-pass, linked in the newsletter) — a candidate tool to trial against Sanity Check's own headline/description pairs before publishing, rather than relying on gut-check.
Related
[[2026-01-07-write-with-ai-headline-claude-skill]] — a different headline-writing framework (Dickie Bush/Nicolas Cole's 5-part structure) encoded as a Claude Skill; same mechanism (headline engineering as a repeatable, testable system) applied to human readers instead of agent readers — worth comparing the two frameworks for where they agree/diverge now that agent-citation has its own data. [[2026-06-04-innermost-loop-bots-pass-humans-on-the-web]] — broader evidence for the underlying shift this piece assumes (agents as a growing share of content "readers"), useful as macro context for why GEO/agent-citation optimization is worth RDCO's attention now rather than speculative. [[project_credibility_for_phdata_sales]] [[feedback_no_derivative_sanity_check_pieces]]