Earlier this month a client wanted to try ads inside ChatGPT. OpenAI had reported 900 million weekly users in February, and some of the client's buyers were among them. The ads went live on a Tuesday. On Wednesday at 11:30 the first click arrived at the site. At 11:31 a quote request came in, and the contact appeared in the CRM with the campaign already attached. At 13:01 a deal existed with an owner assigned, created by an automation reading the form, not by a person. Ninety minutes, and nobody had carried anything between the ad, the site, the analytics and the CRM.
Two things did not happen that day, and they belong in the telling. The ad platform did not record the conversion, because a tag was sending the wrong event, and a person found and fixed that five days later. And the salesperson who owned the deal did not call until the following week. The loop closed in the site and the CRM within the hour. Closing it on the ad platform and with the buyer took people.
Connected, not just faster
Speed is the story most people tell about AI in marketing, and it is true as far as it goes. On the sites we run, a page that took a week takes an afternoon, and a campaign brief takes minutes.
The subtler change is that the site, the campaigns, the CRM and the reporting are now one system. They were not before, and the reason is easy to miss: the joins were people. The site lived with one agency, the ads with another, the CRM with sales, and the report in a spreadsheet someone updated on Fridays. A good part of a marketer's week went to ferrying things between them. Which campaign this lead came from. Whether the page it landed on still said what the campaign promised. Whether the number in the deck matched the number in the dashboard.
How, twice
Two drawings of the same loop.
The first is the idea. Strategy, the Brand Strategy, sets what to say and to whom. The site says it. Campaigns send people to the site. Reporting shows what happened. Strategy is revised from what converted, and the loop goes around again.
The second is how it is built. The site is standard HTML and JavaScript on templates, hosted on Cloudflare with the source in GitHub, so an agent that can read code can read and change it. Every source lands in one data layer: web analytics, search console, the ad platforms, Cloudflare, and the CRM, all in BigQuery. An agent sits at each step of the loop, reading the strategy, editing the site, running the campaigns, writing the report. The operator sits above the loop, and every change goes through them.
Not every site runs every part of this yet: today the strategy and the site are agent-run on all of them, the data layer is filling source by source, and the report is the next piece.
Nothing on the second drawing is new on its own. Templates, hosting, a warehouse, a CRM: all available for a decade. What is new is the agents, and that each of them reads the same strategy and the same data.
Why connected matters
A connected loop learns from its own results. When reporting sits in the same system as the campaigns and the site, what converted last month is an input to what gets built next, not a slide shown once and filed. The strategy is revised, the site is updated to match, the next campaign starts from the revision.
What it learns first is not always what was expected. On one client site this summer, the first thing the reporting step corrected was an inherited number: an authority figure carried over from an old audit did not survive a direct read of the link data. The strategy changed from a content plan to an authority plan the same week. Two cycles of the loop have run on that site so far, and the learning has been about inputs, which keywords, which pages, whether Google had crawled the post at all, rather than about outcomes. The slowest link in the loop is not the reporting. It is Google, which in our experience takes months to settle on a new site. A loop that cycles faster than that re-reads the same state.
It also improves without being rebuilt. The agents at each step are the same kind of thing: a model reading a document and a data layer and proposing a change. When a better model is released, every step gets better at once. The templates, the data layer, and the strategy do not change. This is the part that is hardest to see from outside, because nothing visible happens. The site looks the same. The report has the same columns.
The proposals are better.
What it changes for a small team
For a company with ten to a hundred people, the marketing budget has mostly gone to handoffs: the agency retainer that covers the site, the second one that covers the ads, the hire who keeps the CRM honest, the hours assembling the monthly report. When the joins are not people, that money goes to decisions instead. What to say. Who to say it to. Which of the three things that worked to do more of.
One operator with the loop does what took a team, not because the operator is faster but because the carrying is gone. The closest measured analog is the 2023 study of 758 consultants working with a model: about a quarter faster and 40 percent higher rated quality on the tasks the model was good at, and worse on the tasks it was not.
The ceiling on what a small company can attempt rises with it. A test that used to take a quarter of coordination takes a meeting. Some tests are worth running only because they are cheap to run, and the ChatGPT ads at the top of this piece were one of those: a small budget and one lead in the first week.
What it doesn't change
The loop does not decide what the company is for. Positioning, the choice of who to sell to and what to say to them, is judgment, and the agents write from it rather than produce it. When the strategy is wrong, the loop gets a wrong thing to more people, faster.
An experienced operator or a small team still runs it. In practice that person reads every proposed change, rejects some, and rewrites the ones that are almost right. On the client sites running the loop today, the operator still decides which reviewers to name, which address is the canonical one when the listings disagree, and whether a lead is worth a salesperson's afternoon. Most of what matters in a B2B sale is not in the analytics. A connected loop shows which page converted. It does not show why the prospect trusted the company.
The ninety minutes at the top of this piece ended with a deal record, not a customer. The call that might turn it into one was made by a salesperson the following week. What has changed is where that person's hours go. Fewer of them carry things between systems, and more of them go to deciding what to do next.
Saprel builds the loop: Guide holds the strategy, Web is the site and its reporting, Lifecycle holds the conversations and the pipeline that come back.
Operators run it.