Your Product Copy Was Built to Rank. Agents Read It to Reason.

There are two moments when a description gets read. The first is after a click, when a human who has already expressed interest in an item needs enough information to confirm the decision they're leaning toward. The second is before any click, when an AI agent ingests the same text as raw material to decide whether the listing belongs on a shortlist at all.

Both readers see the same words. The jobs are completely different, and so are the consequences of writing for only one of them.

Most descriptions in the wild were designed for the first moment. Search-engine ranking shaped term density, human reassurance shaped tone, and the call to action assumed someone who arrived willingly, already curious. Nobody was writing for a reader who arrives before the shopper does, without the shopper, to make a gatekeeping judgment on their behalf.

That gatekeeper is here now. And copy that wins at search ranking is often mediocre reasoning material for the system that decides whether a buyer ever reaches your page.

The Agent's Job Before the Click

When a shopper delegates a purchase to an AI agent, it receives a prompt with embedded constraints: a budget, a fit, a use case, sometimes a vibe. The system then queries product feeds, reads what comes back, and decides which items merit a recommendation. Price and availability are structural. Everything else the agent knows about your listing comes from what you wrote.

A 2026 study from Harvard Business Review ran more than 16,000 simulated shopping rounds across four AI models, testing how agents respond to eight common e-commerce persuasion tactics. The headline finding was unambiguous: star ratings were the only signal that pushed selection upward consistently, across all four models and all product categories. Every other badge, including countdown timers, scarcity cues, strike-through pricing, and purchase counts, produced effects that varied by model and category, sometimes dramatically.

The uncomfortable part is what the researchers observed in more capable models. GPT-5 and Gemini 2.5 Pro were less responsive to marketing cues than their lighter counterparts, and in several cases they appeared to penalize overt persuasion, treating heavy sales language as a signal of low quality or manipulation. The researchers put it plainly: the direction of travel is toward agents where more persuasion produces less selection.

That pattern reaches directly into description writing. The same instincts that drive keyword-loaded, conversion-phrase-heavy copy, the superlatives, the urgency language, the aspirational positioning, are exactly the signals that capable reasoning models appear to discount or hold against you. The writing style that served two decades of search and conversion optimization sits in an awkward position with the AI reader.

What Field Data Shows About the Gap

I ran a UCP catalog audit of six Shopify-hosted merchants, sending the same query to each via the same signing protocol. Every merchant returned identical field shapes: title, description, price, variants, tags, images. The protocol is uniform. What ships through it is not.

Across the six merchants, text ranged from 146 characters to 828, across brands in the same apparel niche, on the same protocol, competing for the same potential queries.

Merchant	Chars (html)	Copy approach
Everlane	146	Brand voice tagline, minimal substance
Cuyana	161	Single claim, no differentiators
UNRL	198	Short prose, promotional framing
Taylor Stitch	393	Material and construction callouts
Gymshark	693	Feature-tagged, functional attributes
Oliver Sweeney	828	Fit, fabric, occasion, construction detail

A 146-character description is a tagline. At that length, there is room for a sentence or two of brand voice and a general category claim. An AI reader that ingests it learns almost nothing beyond the title. Everlane's jacket descriptions, at 146 characters, gave AI systems essentially the same information as the product title with slightly different phrasing.

Oliver Sweeney's descriptions, at 828 characters, covered fit (slim, structured shoulder), fabric (full-grain leather, suede), occasion (business casual through formal), and construction details. A system matching against a prompt like "slim leather jacket I can wear to a client dinner" has far more substrate to reason over. The match happens or fails based on what the copy actually says.

Gymshark landed at 693 characters with a different approach: feature-tagged copy that called out breathable, lightweight, reflective, and sweat-wicking in structured terms alongside prose. Their text was optimized for functional matching rather than aspirational positioning.

The merchants with the thinnest copy, 146 to 198 characters, are not facing a protocol problem. The protocol exposes a field that can hold as much text as any merchant wants to write. Thin copy is a content decision. A choice, at least implicitly, about whether the agentic surface is worth competing on.

The Two Failure Modes

The audit revealed two distinct ways merchants lose the agent read.

The first is brevity. A description short enough to fit in a tweet has nothing for a reasoning system to match against. When evaluating whether a listing fits a prompt, the model works with what exists. Absent specifics about fit, material, and use case, it falls back on structural signals like price and whatever it can infer from the title. That is a weak position when a neighboring merchant has written 800 characters of genuine reasoning material.

The second is persuasion density. The HBR research found that capable reasoning systems responded negatively to overt marketing language. Listings loaded with claims about quality, prestige, and desirability may actively reduce selection probability on frontier systems. The problem is compounded by how that writing was typically constructed: keyword-heavy for search engines, phrase-heavy for human psychology, and therefore dense with exactly the signals that capable AI appears to discount.

These failure modes often appear together. A short description usually packs its limited character count with the merchant's highest-priority claims, which tend to be promotional. So the text ends up thin on substance and heavy on persuasion, which is the worst combination for the pre-click read.

What Agent-Readable Descriptions Actually Look Like

The goal is specificity over persuasion. Reasoning material over reassurance. Text written to give a model the facts it needs to make a match, rather than copy written to nudge a human who has already clicked through.

Agent-readable descriptions answer the implicit questions encoded in a shopper's prompt. "Find me a lightweight jacket for a warm-weather hiking trip" encodes constraints about weight, activity type, and climate. The listing that wins that query states directly that the jacket is designed for warm-weather use, gives the fabric weight or technical specification, and says something concrete about fit and packability. The merchant that wrote about "timeless design" and "premium craftsmanship" has given the reasoning system nothing to match against.

Substance, in practice, means fabric and weight in specific terms, alongside whether the fit runs slim, relaxed, or oversized. It means the intended activity or occasion range stated plainly, care requirements if they affect how someone uses the piece, and who the product was designed for. These are the dimensions a model compares against what the shopper asked for, and every detail that's present is a detail that can make the match.

What drops out is language that only functions as persuasion: superlatives without referents, urgency framing without factual basis, aspirational claims about how the wearer will feel rather than what the item does. Gone. Capable AI systems have learned to discount these signals, and in some cases penalize them. Merchants who write as if they're answering a skeptical, capable questioner will fare better as reasoning systems advance than those who spend words nudging a willing buyer.

A useful heuristic: write the description as if you were answering a specific customer service question about the item. "What makes this jacket right for someone who runs warm and needs something for layering?" That question forces specificity. It insists on facts and resists vagueness. The answer to that question is a better description for an AI system than any number of lines about "elevated basics for the modern wardrobe."

The Compound Effect

Descriptions do not change often. A merchant that writes agent-readable copy today writes it once for every SKU. That text then becomes the reasoning material every AI system that queries the catalog uses from that point forward, across every agentic surface that reads the feed, through every model generation that ingests it.

The payoff compounds. The cost of inaction compounds too. Each month that a description sits at 146 characters is another month of queries where the matching fails because there was nothing to match against. Meanwhile, Oliver Sweeney's 828-character description is earning every one of those matches, for every agent query, on every platform that reads the feed, at no incremental cost after the copy was written. That matters more than it once did: AI-referred buyers now convert 42% above baseline and spend 14% more per order, which means the text that wins or loses an agent recommendation is directly correlated with your highest-value acquisition channel.

This post covers the on-page copy itself, the most immediate lever. The broader argument about how authored reasoning, fit guidance, substitution logic, and voice compound across all five dimensions of an agent-ready catalog lives in the brand-truth layer piece. For anyone who wants to see what a fully structured brand-truth artifact looks like in practice, there is a worked example on a real product page.

The listing text is where it starts. Write it for the reader that reasons. That reader is deciding your shortlist position before the shopper ever sees your name.

---

If you want your product copy assessed for agent-readability, or want early access as we build toward the Sartorial beta, get in touch.