“Search” was a single product for the last twenty years. It’s not anymore.
The Google ten-blue-links page that defined what “search” meant from 1998 through 2023 is now one of at least three meaningfully different surfaces. Google itself has split into a page where AI Overviews dominate the top and traditional results are pushed below the fold. ChatGPT, Perplexity, Claude, and Gemini operate as a different kind of surface entirely: a conversation where the answer is a recommendation, not a list. And the walled-garden surfaces (Amazon’s agent, Apple’s Siri commerce path, Walmart’s emerging equivalent) are a third variant, optimized for catalogs the platform owns or partners with.
The brands optimizing for these three surfaces as though they were the same surface are spending against the shrinking one. The brands optimizing for them as three different problems are starting to compound.
This post is the thesis case. Why the surfaces are different, why the differences matter, and what “Agent SEO” actually requires that traditional SEO doesn’t.
The user-side change
The first thing that matters is what the buyer brings to the surface.
A buyer who types “best wireless earbuds for running” into Google brings an implicit acceptance that they will read three to ten results before deciding. The query is the start of a research process. The surface is designed to support that process by ranking links the buyer will navigate.
The same buyer typing the same query into ChatGPT brings a different implicit ask: give me an answer. The surface is designed to support that ask by synthesizing a recommendation from many sources and presenting it as a conclusion. The buyer’s tolerance for “here are some options, you decide” is lower in this surface, because they specifically asked for the synthesis.
This is the structural difference. Google’s value is the rank-ordered set of options. ChatGPT’s value is the answer. The signals each surface rewards follow from this difference.
The ranking dynamics divergence
Google rewards content that earns the click. The signals it uses (backlinks, dwell time, click-through rate from SERPs, freshness, query-doc semantic match) are all proxies for “this page will satisfy the user well enough that they don’t immediately go back to search.”
ChatGPT rewards content that earns the citation. The signals it uses (structured data fidelity, JSON-LD that mirrors prose, honest review distributions, specific numerical claims that can be verified across sources, clean policy disclosure, content that scores high on the agent’s internal “is this reliable” classifier) are all proxies for “we can confidently quote this source in our recommendation without misleading the user.”
The two reward functions overlap but they don’t converge. A page that wins on Google might have impressive backlink authority and an excellent title tag but cherry-picked review widgets and contradictory pricing across the page and the structured data. That page loses on ChatGPT, because the agent catches the contradictions and downranks the source. Conversely, a page that wins on ChatGPT (clean schema, honest review distribution, specific verifiable claims, matched prose and JSON-LD) might still have weak backlinks and rank position 7 on Google.
The brands that have been winning on Google for a decade are surprised when they don’t dominate ChatGPT recommendations. The reason isn’t a bug. The reason is that the ranking functions are different, and the brands that won on Google won partly by gaming the things Google measures. The gaming doesn’t work on the agent surface, because the agent is measuring different things.
The reward function difference
There’s a deeper layer to this. The two surfaces have different definitions of what a “good outcome” is.
Google’s reward function is engagement-shaped. The signals it optimizes for produce pages users click on, scroll through, and stay on. The business model (advertising) rewards engagement directly. The signals propagate down to publishers, who optimize for engagement, sometimes at the cost of accuracy or honest framing.
ChatGPT’s reward function is closer to recommendation-shaped. The signals it optimizes for produce answers users trust, follow, and convert on. The business model is consumer subscription plus enterprise API, and increasingly transaction-attached commerce. Engagement is not the proxy. Trust is. The signals propagate down to publishers (and merchants) in a different direction: honesty becomes more valuable, sleek-but-empty content becomes less valuable, specificity becomes more valuable than rhetorical polish.
The brands that built content strategies for engagement-shaped surfaces have a hard time letting go of the patterns. The agent-shaped surface rewards a different style of writing, a different style of structured data, a different stance on customer reviews, and a different posture toward product limitations. The pattern isn’t subtle once you look for it.
What Agent SEO actually requires
Agent SEO isn’t SEO with extra steps. It’s a discipline with its own deliverables. Here are the six that matter most based on how agents are documented to behave:
Structured data that mirrors prose exactly. Gemini and Claude both penalize mismatches between JSON-LD and visible content. If your Offer block says availability: InStock and the page text says “ships in 3 to 4 weeks,” you lose confidence weight. The fix is to make the structured data the source of truth and have the prose follow.
Honest review distributions. AggregateRating with the real distribution, not a curated subset. Specific written reviews that name real product attributes, including limitations. Pages that surface negatives confidently get cited more, not less. The agents have learned what manicured reviews look like.
Specific, verifiable claims. “Lightweight” loses to “weighs 248 grams.” “Highly rated” loses to “4.6 stars on 1,847 reviews.” “Fast shipping” loses to “ships from California, arrives in 2 to 4 business days for the lower 48.” The pattern is consistent. Agents reward sources they can cite with specificity. They downrank sources that traffic in vague positives.
Policy disclosure in llms.txt. Return policy, shipping cutoffs, warranty terms, restocking fees, gift-return windows. Things buyers actually need to know before purchase. The brands that publish these clearly in their llms.txt get included in agent shortlists more often than brands that bury them in fine print.
Real-time inventory and price truth. Stale stock data is the single most expensive lie a merchant tells an agent. The agent that recommends an out-of-stock product to a buyer learns to distrust that catalog. The agent that recommends a product at last week’s price to a buyer who completes the purchase at this week’s price has a worse outcome to learn from.
Speed-to-answer on /agent/* endpoints. If you’ve published agent-specific GET endpoints (and you should), they need to return in under 200 milliseconds. Gemini specifically times out and switches sources past two seconds. The performance bar is higher than the human-facing storefront.
None of these are revolutionary. All of them are different in tone and emphasis from the traditional SEO playbook. The brands that copy their traditional playbook over to “Agent SEO” by changing the title tag and adding a schema field will be disappointed in the results. The brands that treat it as a different discipline, with its own success metrics, will build a measurable lead.
What you give up
There’s a real tradeoff in optimizing for agents. Some of the patterns that win on agent surfaces are patterns that traditional CRO would advise against. Honest acknowledgment of product limitations reads weaker to a buyer who’s already on your page. Specific numerical claims expose you to scrutiny in a way generic claims don’t. Real-time stock truth means showing “sold out” prominently when it’s true.
We’ve watched this tradeoff play out as the category has matured. The pattern is that the brands that optimize for agent surfaces don’t lose much from the traditional surface, because the traditional surface is shrinking anyway for the queries that matter. The brands that try to keep both perfectly tuned (rich, polished marketing copy for the human visitor, sparse honest copy for the agent) end up with two versions that drift apart and contradict each other, which is the worst outcome.
The simpler path is to write honestly, structure data faithfully, and trust that the same content can serve both surfaces if it’s accurate. The brands that commit to this saw mixed results in 2024 and clear positive results through 2025. The 2026 picture is shaping up the same way.
The acquisition allocation question
The most concrete question this thesis raises is how much of your acquisition budget should sit against each surface.
The wrong way to answer it is to keep allocating against last year’s traffic mix. That mix is shifting under you. The right way is to map your top commercial queries to the surfaces they’re increasingly answered on, attribute the recent visit and conversion trends to the surface mix, and plan against where the surface mix is going, not where it was.
Most brands we’ve talked with overweight Google by a factor of two or three relative to where their 2026 acquisition is actually going to come from in their categories. The correction will be uncomfortable. The brands that make it earlier will be the ones with budget left for testing on the surfaces that matter.
Search is splitting in two. The brands that recognize this in Q1 will allocate their 2026 differently from the brands that don’t. We’ll know by mid-year who allocated correctly.