Remember when Google Reader shut down and a whole generation of publishers suddenly discovered how much of their traffic came from a single piece of software they didn’t control? People had built content strategies around a habit, and then the habit’s host went away. The lesson stuck around longer than the product did: when a machine becomes the primary reader of your work, the machine’s roadmap becomes your roadmap.
That’s the thought I keep circling back to reading Gizmodo’s report that OpenAI is pressing ahead with a new always-on agent. Not a chatbot you open when you have a question. Something that keeps running. For anyone in search, that distinction matters more than most of the feature announcements we’ve been handed in the last two years.
Always-on changes the shape of the traffic
An assistant you summon is bursty. It fetches when you ask, and the pattern of requests roughly maps to human curiosity. An agent that persists is a different animal. It checks. It monitors. It re-reads things it read yesterday to see whether they changed. It doesn’t get tired at 11pm and it doesn’t take weekends off.
If you run a site of any real size, that shows up in places you may not be watching:
- Fetch volume that doesn’t correlate with any human session
- Repeat requests to the same URLs on a cadence, not a whim
- Traffic that reads your content thoroughly and sends back nothing measurable
- Load patterns your caching rules were never tuned for
None of that is inherently bad. Some of it is arguably a compliment. But it does break the mental model most analytics setups are built on, which is that a request implies a person, and a person implies a possible outcome. Agents decouple those. Your content can be doing its job perfectly and your dashboard can look flat.
What I’m actually telling clients
Start logging with the assumption that a meaningful slice of your readers are software with a persistent memory. Practically, that means separating agent and human traffic in reporting rather than lumping it into a “bot” bucket you ignore. It means checking whether your rate limits punish a well-behaved agent that comes back hourly. And it means making sure the version of your page a machine reads is the version you’d want quoted, because with an always-on system, the thing that gets cached could get referenced for a long time.
Structured data earns its keep here. So does having your key claims in text rather than in an image or a chart with no caption. Solid internal linking helps a persistent reader understand which pages you consider authoritative. This is not new advice. The difference is the audit frequency changed from occasional to constant.
The rest of the week’s news is part of the same story
Gizmodo also reported that OpenAI’s head of data centers has apparently left, and that the company’s CFO reportedly downplayed the importance of an IPO to staff, framing it as “another fundraise.” Meanwhile Sam Altman has conceded it’s going to take a long time before AI goes mainstream.
Put those next to the always-on agent and you get a picture that’s less triumphant and more human. Persistent agents are expensive. They consume compute continuously by design, which makes data center leadership a load-bearing role, and makes fundraising a recurring event rather than a milestone. Altman’s mainstream comment reads to me like expectation management from someone who has looked at both the adoption curve and the bill.
For search people, that’s useful context. It suggests agent traffic will grow, but not overnight, and not evenly. You have time to prepare. You don’t have time to ignore it.
The part that gives me pause
Gizmodo’s writer noted they usually laugh off AI hacking reports, but found a recent one genuinely serious. I don’t have the details to analyze, and I’m not going to pretend otherwise. What I will say is that “always-on” and “security concern” are two phrases that get more uncomfortable the closer they sit together. A system that runs continuously with access to accounts and browsing has a wider exposed surface than one that wakes up on request.
If you’re integrating agents into your own workflows, and plenty of SEO teams already are for auditing, monitoring, and content drafting, scope the permissions tightly. Give them read access where read access is enough. Keep credentials out of anything that runs unattended.
Where this leaves us
I’m not calling this a turning point. Persistent agents are a reasonable next step for a company that has been building toward them openly, arriving alongside signs the underlying economics are still being worked out.
What I’d do this quarter is unglamorous. Look at your logs. Learn which agents visit and how often. Decide deliberately what you serve them. The publishers who got burned by Reader weren’t wrong to build for it. They just never checked how much of their reach depended on it.
🕒 Published: