AI SEO Is Splitting Into Retrieval, Citation, and Agent Execution

Search ranking is beginning to learn from agent behavior

Learning to Retrieve from Agent Trajectories gives us unusually direct evidence. The researchers completed 34,619 search-agent runs, retained 26,482 correct trajectories, and converted them into 91,713 training pairs. Their LRAT method learned from which results an agent browsed, which candidates it left unbrowsed, and whether the retrieved material appeared in its immediate reasoning.

Across six tested agent backbones, the trained retrievers improved evidence recall and usually improved final-task success while using fewer interaction steps. One out-of-domain configuration tied rather than improved. That detail matters because this is promising evidence, not a universal claim that every agent and every retrieval system will behave the same way.

The larger change is the feedback signal. Traditional search systems have long used forms of human interaction and relevance judgment. LRAT shows a working way to train retrieval from the trajectory of an agent trying to finish a task. The retriever is being rewarded for evidence that helps the agent reason and continue, not merely for a result that looks relevant to the original query.

• The positive signal came from results the agent chose to browse inside successful trajectories.
• Hard negatives included candidates the agent saw but did not browse.
• The paper reports evidence recall, task success, and interaction steps, not public Google rankings or business conversions.

The next query may come from the evidence on your page

A separate SIGIR study analyzed 14.44 million requests sent to the DeepResearchGym search API between June and December 2025. After the researchers grouped requests into sessions, 90 percent of multi-turn sessions contained no more than ten requests. Fact-seeking sessions increasingly repeated near-duplicate queries, while procedural sessions showed more refinement.

The number that deserves close attention is 54.35 percent. Across multi-turn sessions, that share of newly introduced query terms exactly overlapped with evidence that could be retrieved from earlier steps. This demonstrates lexical traceability. It does not prove that the earlier evidence caused the new query or improved the final answer, because the study did not retain original result pages or score answer quality.

My practical conclusion is still important. Query fan-out can become an evidence-feedback loop. The original prompt may start the process, but terminology, entities, attributes, constraints, and facts encountered during retrieval can shape what the agent searches for next.

This changes what useful writing looks like. A page should contain stable evidence units that can survive retrieval and reuse: exact entity names, explicit relationships, numerical facts, definitions, product attributes, dates, constraints, and sourceable claims. Terminology should stay consistent across related pages. I am not claiming that traditional internal-link optimization has been proven to improve agent trajectories. Neither paper tested that. I am saying that vague copy gives a reasoning system less precise material to carry into its next search.

The GEO evidence is much narrower than the industry makes it sound

A July 15 critical survey reviewed 45 core GEO studies published from late 2023 through July 2026. Its most valuable conclusion is not a new optimization trick. It is the boundary around what current research can honestly prove.

There is reasonable evidence that content already supplied to a model can affect citation and answer composition. Topical relevance and a source's position inside the supplied context are among the most reproducible factors. Evidence for a durable method that improves organic discovery across public AI-search products is much weaker. Evidence connecting those changes to business outcomes is weaker again.

Fixed rewriting formulas transfer poorly between systems and conditions. Structured, extractable evidence has more support than generic style changes, but the benefit depends on the task and retrieval setup. Competition can dilute a tactic as other publishers adopt it. In at least one controlled pipeline, citation-oriented rewriting also damaged retrieval performance. None of this supports selling a universal GEO checklist as a ranking formula.

The review is itself a single-author arXiv preprint and a critical scoping review, not a meta-analysis. That limitation belongs beside its conclusions. Still, its separation of the causal chain is exactly right: activation, crawling and indexing, retrieval, reranking and context allocation, generation and citation, factual absorption, attention, click, and conversion are not interchangeable events.

• A page can influence an answer without receiving a visible citation.
• A citation can appear without producing a click or a business result.
• A retrieved page can lose during reranking, context allocation, or answer construction.
• A formatting change that helps one engine or prompt family may fail when the engine, competition, or context changes.

One AI visibility score is structurally wrong

This is why I do not trust a single score that claims to summarize AI-search performance. It combines unlike events, hides missing data, and makes it impossible to identify the point of failure. A citation percentage is useful if the question is citation recurrence under a documented prompt set. It is not a complete ranking system, and it is not revenue attribution.

A serious test should disclose the engine and mode, market, date, prompt family, paraphrases, run count, source eligibility, control pages, context order when controlled, competitor conditions, and human verification method. Public answer systems change, so the same test should be repeated in separate time windows. A before-and-after screenshot from one prompt is an anecdote, not an experiment.

The reporting should remain staged. If crawl eligibility improves but retrieval does not, the next action is different from a case where retrieval improves and factual absorption fails. The purpose of measurement is to tell us what to do next, not to create an impressive number for a dashboard.

Being found is not the same as being usable by an agent

The July 13 agent-ready website experiment tested this distinction directly. Researchers built two versions of the same local e-commerce site and ran 300 independent browser-agent sessions across GPT-4.1, Gemini 2.5 Flash, and Grok 4 Fast. The agent-ready version retained JavaScript but also exposed product facts through JSON and JSON-LD, explicit identifiers, stock and availability fields, clear control labels, evidence pages, and temporal-validity information.

Strict task success rose from 49.3 percent, or 74 successful sessions out of 150, to 89.3 percent, or 134 out of 150. Average interaction steps fell from 9.31 to 6.49, a reported reduction of 30.4 percent. The largest statistically significant gains were in product-detail extraction and product comparison.

That is a large controlled result, but it is not production-web proof. The test used one synthetic local store, bundled several improvements together, and did not isolate which feature caused the gain. It did not measure crawling, indexing, ranking, public AI citations, referrals, conversions, or real shoppers. The right conclusion is that the pattern deserves paired task testing on staging, especially for commerce, bookings, lead forms, local services, and comparison experiences.

This creates a discipline beside SEO and GEO. A page can rank and be cited while remaining difficult for an agent to operate. The machine may not be able to identify the correct product, confirm that it is in stock, distinguish a primary action from a decorative control, or know whether a claim is still current. Agent readiness starts where visibility ends.

Schema can improve understanding without improving citation visibility

This helps explain an apparent contradiction in the schema discussion. Ahrefs first found a strong correlation between JSON-LD and pages already cited by AI systems. It then followed a cohort of 1,885 already visible URLs that added schema and compared them with a matched control group. AI Mode and ChatGPT citation changes were statistically indistinguishable from zero. AI Overview citations declined slightly relative to controls, but the study could not establish that schema caused the decline.

The study does not tell us what schema might do for previously uncited pages, individual schema types, JavaScript-injected markup, or longer time horizons. It does tell us not to confuse correlation with intervention. Pages that earn citations may be better maintained and more likely to use structured data, without schema being the reason they were selected.

My view is that schema still matters, but the claim must be precise. Structured data can give a crawler or agent a cleaner representation of identity, price, availability, dates, relationships, and actions. That can reduce ambiguity during extraction or execution. It does not follow that adding markup will cause a public answer engine to retrieve or cite the page more often.

The large AI-content study points to a quality problem, not a binary switch

Ahrefs also published a July 27 observational study across separate ranking, indexation, and Search Console samples. In the sampled top-three results, its detector classified 5.3 percent as 100 percent AI content. URLs estimated at 80 percent or more AI-generated appeared throughout the top ten. The indexation proxy fell from 49.28 percent for the low-AI group to 40.35 percent for the very-high-AI group. Low and moderate groups also received roughly two to three times more impressions than the highest groups.

This does not prove that Google rewards AI content, ignores it, or applies no penalties. The detector is probabilistic, only evaluated pages with enough text, and is not Google's system. The indexation measure was a proxy assembled from keyword presence, Search Console impressions, and exact-site searches rather than URL Inspection data. The analysis is correlational, so it cannot separate AI use from originality, site quality, topic selection, authority, or publishing scale.

What it weakens is the simple theory that detectable AI use creates an automatic exclusion line. The more credible operational problem is selection and quality. Large automated systems can publish generic, repetitive, poorly sourced material faster than editors can verify it. That material may underperform because it contributes little, not merely because software helped write it.

Trying to humanize copy to fool a detector is the wrong objective. The useful objectives are information gain, original evidence, source quality, factual accuracy, clear judgment, maintenance, and portfolio-level controls that stop weak material before publication.

We can now test retrieval and reranking in a controlled lab

A SIGIR 2026 demo paper presented MCP servers for Pyserini and RankLLM. Together they expose sparse, dense, learned-sparse, and multimodal retrieval, prebuilt or custom indexes, LLM reranking, and standard relevance evaluation to compatible agents.

This does not reproduce Google, ChatGPT, Perplexity, or another proprietary answer system. The demo paper also did not introduce new retrieval algorithms or separately prove effectiveness beyond the underlying toolkits. What it gives us is a practical laboratory where content variants can be indexed under controlled conditions, retrieved for documented queries, reranked, and evaluated against explicit relevance judgments.

That is far more useful than trying to reverse-engineer an unstable public product from isolated screenshots. We can test whether a definition is retrievable, whether terminology changes candidate recall, whether a comparison block survives chunking, and how a reranker orders evidence. To measure whether the evidence improves the final answer, we still need a separate downstream evaluation with answer-quality judgments.

My read: AI SEO is becoming an evidence and execution discipline

The most important change is not a new acronym. It is that the systems between a question and an outcome are becoming easier to separate and test.

For years, SEO could focus heavily on whether a person saw a result and clicked it. The new environment includes systems that search repeatedly, borrow terminology from retrieved evidence, combine claims from several sources, and sometimes act on the website without a traditional browsing journey. A page now has to be useful to retrieval, reliable during synthesis, understandable to a machine, and safe to operate.

I believe the websites that win will not be the ones that produce the most machine-polished copy. They will be the ones that publish facts worth retrieving, keep those facts consistent, expose the right structure, show where claims came from, state when information is valid, and make important actions unambiguous.

This also changes the work I would prioritize. First, replace the single AI visibility score with staged measurement. Next, build agent-ready templates for the pages where a machine may compare, select, book, buy, or submit. Then evaluate content systems on factual contribution and retrieval utility, not on whether a detector thinks the prose sounds human.

None of this makes conventional SEO irrelevant. Crawlability, canonicalization, page speed, clear information architecture, useful internal relationships, and trustworthy content remain the foundation. The difference is that ranking and clicking are no longer the final tests. We also need to know whether the evidence survives the answer and whether the website works when an agent arrives.