Running an AEO programme

An AEO programme is the loop that turns the earlier chapters into routine work. Each cycle starts from the measurement, and the measurement decides what the team works on next.

The difference from a one-off audit is that the backlog never comes from opinion. It comes from the gaps in the latest answers, each routed to the team that can close it.

What does each cycle do?

1. Measurement. The prompt set runs on the tracked engines, scoped to what the brand sells, as chapter 3 sets out.
2. Reading. Each gap is placed by stage, using the ledger in chapter 1, and by who the answers cite, using the sort in chapter 5.
3. Routing. The gap type names the team that owns the lever, as in the table below.
4. Action. Each team works its own list.
5. The next measurement. Results are judged per engine against the baseline, on the questions the work was meant to change.

A gap that fits no row is usually a scoping question: a question outside the brand's market belongs out of the prompt set, not on anyone's list.

What does the first cycle look like?

The first cycle is the baseline, and it starts with access. If an engine cannot reach the site, every number for that engine describes the block rather than the brand, so the four layers in chapter 4 are checked before the first reading is trusted.

The first cycle also settles the prompt set and the competitor list. Both stay fixed afterwards, so the time spent getting them right is repaid in every later comparison.

What does each cycle's report contain?

A short, repeatable report keeps the loop honest. Each cycle's report carries:

• The four signals per engine, against the baseline and against the usual movement between runs.
• The gap list by owner, with each gap's type and the questions it covers.
• Last cycle's actions, each with the questions it was meant to change and whether those questions moved.
• Any change to scope, such as a new product line added to the prompt set, with its own baseline.

The third item is the one most reports leave out. Without it, the programme records activity but never learns which work moves the answers.

How often should the loop run?

The measurement runs continuously, and decisions are made per period. A team reads the answers on a fixed rhythm, weekly or monthly, and compares periods rather than single runs.

The prompt set changes rarely. When it does, the change is recorded and the new set gets its own baseline, so that a change in the numbers is never a change in the questions.

Should the work be in-house or with an agency?

The split follows the levers. Access work needs control of the site, its firewall, and its templates, so it sits most easily with the team that runs the site. Content and earned work can sit with an agency as well as in-house, since both depend on craft more than on system access.

The measurement belongs with whoever is accountable for the result. When an agency does part of the work, both sides read the same prompt set and the same signals, so neither is grading its own work on a separate scale. Agencies running several brands need the same discipline per client, which AI visibility platform for agencies managing multiple client brands covers.

How should a measurement tool be chosen?

A tool is chosen against the method in this guide: scoped prompt sets, answers as the unit, the four signals reported separately, and per-engine reading. How to evaluate AI visibility tools as answers shift walks through a trial, and the evaluation checklist turns it into a form.

What do AI agents and MCP change?

An AI agent pursues a goal through several steps, calling tools along the way. Agentic search is the search form of it: the agent plans several searches and reads sources across steps before it answers.

For a brand, most of this runs through ground the guide has already covered. Agents fetch pages on request through crawlers like those in chapter 4, so the same access layers decide whether they can read a site.

The Model Context Protocol is an open-source standard for connecting AI applications to external systems, supported by assistants including Claude and ChatGPT. It lets an assistant use a tool or data source directly. For most brands it is worth watching rather than acting on: the questions buyers ask still reach the answer through the stages in chapter 1.

Where does the guide go from here?

The seven chapters describe one loop: how answers are built, where buyers meet them, how to measure them, and the three levers that change them. The glossary holds every term used along the way.

Zumi runs the measurement side of the programme: the prompt set, the four signals per engine, and the gaps that show each team where to work. Book a demo