You open Perplexity and ask it to “recommend a Taiwanese vendor that makes cleanroom equipment.” It doesn’t answer in one line. It searches, clicks a link, hops from that page’s body to another page, works through the official site, a review, a forum thread — five or six pages of evidence — and only then speaks.

What it just walked through wasn’t your one carefully-optimized product page. It was your whole site.

“Answer once” is becoming “send an agent to walk around”

A year ago, AI search was simpler: hand the model your question, it picks a few sentences from a handful of retrieved results and stitches an answer. What the major engines do now is agentic search — they dispatch an agent that decides for itself what to search, which link to click, whether to jump from this page to the next, and when it has enough evidence to stop.

That shift swaps out the entire GEO battlefield. What you used to optimize was “which page a query lands on.” What you now face is “how an agent moves once it’s inside your site.” Its path is live: what it reads on this page decides what it searches next and where it clicks.

Single-page optimization wobbles once an agent walks the whole site

A May 2026 study from ShanghaiTech (arXiv:2605.12887) put this to the test, and the finding stings for a lot of people: page-level GEO tactics — rewriting and polishing a single page — did not reliably beat a site that wasn’t optimized at all under agentic, multi-round search.

This isn’t to say page optimization does nothing. It’s that page optimization can’t govern where the agent goes next. The researchers’ explanation is blunt: page-level moves only touch a single page’s snippet or summary; they can’t shape the agent’s overall browsing trajectory. You buff one page to a shine, and the agent may read it and follow one of its links straight off into a corner you never touched — or come in through a different door and never pass your good page at all.

Single-page work decides “is this page good.” It can’t decide “will the agent reach this page, and once it does, where does it go next.” In a world where the agent walks the whole site, that second question is what settles the match.

What ecosystem-level optimization actually looks like

The same study proposes a method called TRACE, whose idea is to promote the unit of optimization from “one page” to “one connected evidence ecosystem.” It looks like this:

  • One navigation entry page, acting as the hub the agent enters through;
  • Six supporting pages in different roles — official explainer, third-party review, expert take, news coverage, forum discussion, social content;
  • All of them tied together by shared terminology, consistent product attributes, and internal links.

The gap is not small. TRACE’s final recommendation rate landed between 67.2% and 73.9%, ahead of the second-best page-level method by 14.9 to 31.3 points (the benchmark used 3,124 query–fictional-product pairs; the products were deliberately made up so the model couldn’t answer from memory).

The part worth remembering is what the ablation split apart — two independent sources of gain:

  1. “Multi-page evidence that’s coordinated” all by itself — even without a navigation entry page, as long as a few pages agree with and connect to each other, the recommendation rate already beats a heap of unconnected, every-page-for-itself pages.
  2. On top of that, adding a navigation entry page lifts it further still — and it sharply raised how often the agent crawled along internal links.

In plain terms: making your pages know each other is the first layer of gain; giving them a shared front door is the second.

The most counterintuitive point in this study, and the one to hold onto: only a method like TRACE lets the agent crawl to more supporting pages through the links embedded inside a page — something you cannot get from links a search engine hands back.

Think through the difference. What a search engine returns to the agent is a flat list of candidates that have nothing to do with each other. But when the agent clicks into one of your pages and finds a link in the body leading to another of your pages, it now has a browsing channel that sits outside the ranking mechanism. Your internal links effectively draw a few extra roads on the agent’s map — roads that lead only to you.

And that navigation entry page does a second job: it raises the agent’s first-crawl rate on you, and it shapes how the agent reformulates its next query — being read a step earlier changes what it searches afterward. Whether those linked-together pages should sit under one domain or be scattered across subdomains directly affects how your signals concentrate; we break that down in subdomain vs subdirectory vs multi-site.

There’s an expensive illusion this study also punctures in passing: simply getting the same page seen by the agent over and over (delayed crawls, repeated exposure) does not raise the final recommendation rate. Exposure isn’t trust, and it certainly isn’t getting written into the answer. To actually land in the agent’s recommendation, you need a cross-page, coordinated evidence network — not one page working the room.

That cuts down a common calculation: the idea that “lay down a few more links, publish a few more posts, get the AI to see me a few more times” earns a recommendation. Being seen and being chosen are two different things.

What “consistent” actually refers to

The core word in an evidence network is “consistent.” When an agent stitches across pages, it runs into three levels of agreement — or contradiction:

  • Consistent terminology: does the same product, the same service, go by the same name on every page, or three names across three pages?
  • Consistent facts and attributes: do the specs, the numbers, the company details line up across pages, or fight each other?
  • Consistent structure: does every page use a heading hierarchy and navigation an agent can read at a glance to judge “what is this page about, is it worth crawling deeper?”

That third one — structure — often gets filed under aesthetics, when it’s actually an independent, optimizable dimension. A March 2026 study from the University of Tokyo group across six AI engines (arXiv:2603.29979) found that changing structure alone, without touching the content, lifted citation rate by 17.3% (statistically highly significant), holding across all six engines.

And the three layers of structure carry very different weight. Their ablation showed: macro-structure (heading hierarchy, navigation, document flow) contributed 44.9% of the total gain; meso-structure (paragraph organization, lists, table chunking) 39.7%; micro-structure (bold, italic, that kind of emphasis) only 15.4%.

Which means: the agent finds its way by your heading hierarchy and navigation, not by which words you bolded. The heading hierarchy is the agent’s table of contents; getting navigation and levels right does far more than sprinkling bold through the body. How cross-page template sameness gets you downranked in return is unpacked further in site-wide cross-page consistency.

There’s one more gap to be clear-eyed about: getting picked up by the agent and getting written into the answer are two separate gates. Another study dissecting citation behavior (arXiv:2604.25707) splits them into “selection” and “absorption” — a page first has to qualify to be chosen as a source, and then it has to carry evidence that can be lifted directly (numbers, definitions, comparisons, step-by-step procedures) before it’s actually absorbed into the answer. In that same measurement, content with figures and statistics absorbed about 60% better than baseline, content with definitions nearly 60% better, while plain Q&A formatting slightly hurt. So the point of an “evidence network” isn’t only the network — it’s the evidence: every node has to be able to carry a fact worth citing.

A pile of pages, or one evidence network

Put the two mindsets side by side and almost every cell is the opposite:

A pile of pages that don’t know each other One evidence network
Unit of optimization The single page The relationships between pages
Between pages Every page for itself, no links Shared terms + internal links + consistent attributes
Effect on the agent Only touches a snippet / summary Shapes the whole browsing trajectory
Entry Wherever the search engine happens to drop it Navigation entry page as the hub
Exposure logic Rack up a few more impressions Coordinated across pages, then it converts to a recommendation
Under agentic search Doesn’t reliably beat no optimization Recommendation rate 67–74%, ahead by 15–31 points

The verdict is plain: in the agentic era, your site has to be one evidence network. An entry page, internal links, and terminology and facts that agree across pages are the ticket to being reliably recommended in the next stage — not a bonus, the ticket.

So this isn’t something you patch page by page

It’s tempting to reach straight for the tools here: I’ll unify terminology on every page, add an entry page, wire up the internal links, done. Right direction — and exactly where it’s easiest to fix one thing and break another. You change the product name on page A and forget page B still runs the old spelling; you add the entry page but never notice the spec numbers on three pages stopped matching long ago. The difficulty of cross-page consistency is that it’s cross-page: standing on any single page, you can’t see the contradiction; it only surfaces when you lay the whole site out side by side.

Which is precisely the blind spot a single-page checkup can’t reach. A page can score 90 on its own while the whole site together comes out at 60 — because what’s costing you points is the relationship between pages, not the content of any single one. That gap you can’t see on one page but that surfaces across the site — the site-wide GEO audit piece lays out four of these blind spots.

Things you can do this week:

  1. Take stock of the entry. Is there a page on your site an agent (or a person) can start from and, following links, reach all the related explainers, evidence, and cases? If your important pages are islands, connect that path first.
  2. Reconcile three facts. Pick three key facts — company/product name, one core spec, a number you cite often — pull up every page that mentions them, and check whether the wording agrees. This step alone startles a lot of people.
  3. Don’t treat it as a one-off you hand-patch to completion. The agent walks a living, changing site; what you have to maintain is ongoing cross-page consistency, not a single pass and done. The gaps are usually cross-page, cross-system engineering — single-page optimization can’t close them, and eyeballing page by page can’t hold them.

To see whether your whole site reads, to an agent, as one evidence network or a pile of pages that don’t know each other, run a site-wide analysis on geoweb.tw — it lays out cross-page terminology, structure, and entity consistency together, not just a single-page score. Once you can see the gaps and want to know how to close them and which pages and systems they touch, send the URL or the report to [email protected] and we’ll come back with a specific read.

(One caveat: the numbers above come from a controlled benchmark, using fictional products and a specific AI-search-agent setup; against real engines and real brands, the absolute figures will differ. What doesn’t change is the direction — the agent walks the whole site, and a coordinated evidence network holds up better than an isolated, pretty page.)

Further reading