The line between "search engine" and "coworker" just got thinner. Perplexity has introduced Teammates — always-on AI agents that take on roles inside a workspace, inheriting projects, skills, and memory from the Perplexity tools a user already relies on. They plug into Perplexity Computer, a system that can deploy agents on a user's own device. The product details are worth reading directly at the source.
For search and content teams, this is not a product-news story. It is a distribution story. When the thing that used to answer questions starts taking actions, the demand side of the web changes shape: fewer clicks, more delegated tasks, and a new class of reader that is neither a human nor a classic crawler. That has direct consequences for indexing strategy, structured content, and how you earn AI citations.
From Answer Engines to Acting Agents

Figure: Editorial context image from the source page (preview only; rights remain with the publisher) · Credit / further reading: www.theverge.com
The last two years of generative search trained publishers to optimize for being quoted. Retrieval-augmented answer engines reward pages that state facts cleanly, name entities unambiguously, and survive being summarized in two sentences. Agentic systems raise the bar: they reward being usable.
There is a meaningful difference between a system that cites you and a system that executes against your content. A citation is a mention. An execution is a dependency — the agent pulls your data, applies your documented process, or calls your endpoint, and then reports back to a human who never saw your page. The content that wins in that world is not the most persuasive prose; it is the most unambiguous, well-labeled, and self-describing material on the topic.
The Unit of Citation Is Shifting From Page to Fact
Classic SEO reasoned in documents: one URL, one intent, one ranking. Generative engines already broke that into passages. Agents break it further into assertions, parameters, and steps.
Consider what an agent needs in order to do something with your content rather than merely repeat it:
- A declarative claim it can verify — a figure, a definition, a compatibility statement.
- A stable identifier — a product name, a version number, a policy code — that survives paraphrase.
- A scope boundary — what the claim does not cover, which agents need in order to avoid over-generalizing.
- A retrieval path — a canonical URL plus machine hints so the agent can re-check the source rather than trusting a cached snippet.
Most editorial pages supply the first item and neglect the other three. That is the gap GEO has to close.
The Index Is No Longer Only for Readers
When agents inherit memory and act on a schedule, they re-query. Not once at discovery, but repeatedly, at the moment of execution. That makes freshness and machine legibility compounding assets rather than one-off wins.
llms.txt and the emerging machine contract
An llms.txt file at the domain root is the simplest version of a machine contract: a curated map telling language models which pages matter, what they cover, and where the canonical versions live. It is not a ranking mechanism and it does not replace crawling, but it reduces the ambiguity an agent faces when it must decide which of your forty blog posts is authoritative on a given question. For brands with sprawling documentation, product catalogs, or policy pages, this is cheap to ship and disproportionately useful.
Structured data as the agent API you already have
Schema.org markup was built for rich results, but it reads like an API to an agent. Product, FAQPage, HowTo, Organization, and Article with clear author, dateModified, and citation fields give a machine the scaffolding to trust and reuse a passage. The practical rule: if a human would need a table to understand it, an agent needs structured data to act on it. Publish the table and the markup.
Memory Is the New Ranking Surface

Figure: HEXOPS original illustration inspired by the topic (not a copy of the source article imagery) · Credit / further reading: www.theverge.com
The most under-discussed detail in the Teammates framing is memory carried across sessions and roles. Persistent memory changes the economics of being cited.
In a stateless search world, every query is a fresh competition — you can win today and lose tomorrow. In a memory-bearing agent world, the first source that answers well tends to become the default source. The agent remembers the pattern, not just the result. Compounding advantage goes to whoever is correct first and correct consistently, which is a very different optimization target than ranking for a head term.
This has three concrete implications:
- Consistency beats volume. Ten pages that contradict each other on a spec will be discarded by an agent faster than a single well-maintained page. Audit for internal contradictions before publishing more.
- Stable URLs beat campaign URLs. If your canonical answer moves every quarter for tracking reasons, you are resetting the memory you worked to build.
- Attribution must be legible. Agents summarizing on your behalf will increasingly be asked where a claim came from. Clear bylines, dates, versioned updates, and explicit sourcing make your content the easier citation.
A Practical GEO Playbook for Agent Traffic
- Audit for extractability. Take your top twenty pages and ask: could a competent agent lift a correct, self-contained answer from each without surrounding context? If not, rewrite the opening of each section.
- Ship an llms.txt. Curate it by intent, not by date. Point to canonicals, not to every variant.
- Close the structured-data gap. Prioritize
Organization,Product,FAQPage, andArticlewithdateModified. AddsameAslinks so entity resolution is trivial. - Consolidate cannibalized answers. Agents do not enjoy choice paralysis. One authoritative page beats three competing ones.
- Log agent traffic separately. Identify AI-agent user agents and referrers in your analytics so you can measure delegated sessions — the visits where an agent, not a person, did the reading.
- Publish machine-readable policies. Return windows, compatibility matrices, and limits belong in tables and markup, not buried in paragraph six.
What to Watch Next
The competitive question is no longer whether agents will mediate work; it is whether they will mediate your category with or without your content in the loop. Perplexity is positioning Teammates as role-based and workspace-native, which suggests buyers will evaluate agents the way they evaluate new hires — on whether the agent knows the domain. Domain knowledge is largely a retrieval problem, and retrieval is a content-formatting problem. That is squarely inside the remit of an editorial team.
Key takeaways
- Perplexity's Teammates signal a shift from answer engines to acting agents that inherit memory, skills, and project context.
- Agents reward usable, structured, unambiguous content more than persuasive prose; the unit of citation shrinks from page to fact.
- llms.txt, schema markup, and stable canonical URLs function as a lightweight machine contract that reduces retrieval ambiguity.
- Persistent agent memory creates compounding advantage for sources that are correct first and consistent over time.
- Measure agent-mediated sessions separately so GEO work has a feedback loop, not just a theory.
Next steps
- Run an extractability audit on your twenty highest-value pages this week.
- Draft and publish an llms.txt curated by intent and canonical URL.
- Add
dateModifiedandsameAsto your core entity pages. - Fix the top five internal contradictions in product or policy content.
- Instrument AI-agent traffic in analytics and set a baseline before you optimize.
Sources
- The Verge AI — Perplexity launches its own take on AI agents for work



