A shopping agent built on a large model doesn't reason from a blank slate. It carries an opinion, formed from training data and reinforced by live retrieval, before it ever looks at your product page. If that opinion is thin, outdated, or shaped by a competitor's louder presence, a technically perfect page doesn't fix it.
That's the gap most "AI SEO" advice skips. Structured data and readability get you found. They don't get you recommended.
Two separate problems, one gets all the attention
The first problem is technical: can the agent parse your page at all. Clean JSON-LD, a readable page structure, no JavaScript-only rendering blocking the content. This is genuinely worth doing and genuinely cheap.
The second problem is reputational: does the agent already have an opinion about you, separate from your site entirely. This one is harder, and it's the one that actually decides whether you get recommended or your competitor does.
Worth calibrating expectations on the first problem specifically. Research on llms.txt, the proposed standard for telling AI crawlers what's on your site, found it's barely used by major crawlers in practice, a few hundred hits in a study of half a billion requests. Ship it, it's cheap, but treat it as insurance, not the lever that moves outcomes.
Where the actual opinion comes from
Two sources feed it. Training data: whatever the model absorbed about your brand during training, which for most brands means whatever got written about them in reviews, forums, and comparison content, not their own marketing copy. And live retrieval: when an agent searches the web in real time, it's pulling from the same kind of third-party sources, weighted by how much it trusts them.
Reddit specifically carries outsized weight here. One analysis found it accounts for the largest share of citations in Perplexity's answers, and it shows up heavily in Google's AI Overviews too. A brand that's never discussed there, or only discussed by its own marketing team, is invisible to exactly the layer that's forming the agent's opinion.
A shopping agent can recommend your competitor even with perfect technical SEO
This is the scenario worth sitting with. Your JSON-LD is clean, your llms.txt is shipped, your product pages load fast and parse cleanly. None of that stops an agent from recommending a competitor, if the sentiment it's pulling from elsewhere favors them. The technical fix and the reputational reality are almost entirely separate systems, and most teams only work on the one they can control from inside their own codebase.
The uncomfortable part is that the reputational half is largely made of conversations that already happened, in places a brand doesn't own and often isn't watching. Reviews, Reddit threads, forum comparisons, half of them without the brand ever tagged or notified.
What to actually check
- Ask the agents directly. Open ChatGPT, Claude, or Perplexity and ask what they know about your brand and how it compares to two named competitors. A thin, generic, or wrong answer is itself the signal, not a prompting problem to work around.
- Check where you're actually discussed. Search your brand name on Reddit and see what comes up. If the answer is nothing, or only your own posts, that's the gap.
- Read the comparison threads, not just your own reviews. The conversations that shape an agent's opinion are often "X vs Y" threads where nobody tagged either brand. Those rarely show up in a standard review-monitoring setup.
Fixing the technical half takes an afternoon. Fixing the reputational half means actually knowing what's being said, continuously, in the places that feed these models, not once before a launch. That's the specific gap we built Hugo to close: it reads reviews, Reddit, and forums continuously and surfaces what's actually being said about a brand, sourced, instead of leaving that half of the picture to a one-time audit.
Perfect schema gets an agent to your page. It doesn't get the agent to trust what it finds there over what it already believes.