The argument · Worked example
One question, followed all the way through
The clearest way to explain why the parts belong together is to take one ordinary question and follow it until something has actually changed.
20 August 2026 · 7 minute read
Take a question a growth-stage consumer brand actually gets asked: a competitor has started appearing in AI assistant answers ahead of us, why, and what do we do? Answering it properly requires knowing which competitor set you are measuring against, what that competitor published recently, what customers are saying about both of you, which of your customer groups the affected questions matter to, writing something in your own voice that closes the gap, publishing it, running support behind it, and then finding out whether it worked. That is seven steps. In most stacks at least four of them cross a tool boundary, and each crossing is where the work stalls, gets re-entered by hand, or quietly stops.
The question
It usually arrives informally. Somebody on the team asked an assistant what to buy in the category, saw two rivals named and not their own brand, and put it in a message. It is a good question and it is genuinely urgent, because unlike a search ranking there is no second page to be on.
What answering it actually takes
| Step | What it needs | Where it usually breaks |
|---|---|---|
| 1. Confirm it is real, not one bad prompt | The same question asked repeatedly, across engines, over time | A one-off screenshot becomes a strategy meeting |
| 2. Identify who is being named instead | A competitor set that matches the one used everywhere else | The rival is not in the tracked set, so nobody has history on them |
| 3. Find out what changed | What that competitor published, and what got cited | Lives in a different tool, or nowhere |
| 4. Check whether customers agree | What people are actually saying about both brands | Listening tool has a different competitor set and its own relevance rules |
| 5. Decide whether it matters | Which of your customer groups those questions come from | Segments live in the commerce or CRM tool, not the research one |
| 6. Write and publish the answer | Content in your voice, structured to be quotable | Briefed out, comes back generic, takes three weeks |
| 7. Find out whether it worked | The same measurement as step one, plus what the campaign did | Nobody goes back and checks |
Step seven is the one that disappears most reliably. Not through negligence: by the time it is due, the team is on the next thing, and checking requires opening the tool from step one again with a question nobody has written down.
The same seven steps when the parts share context
- Steps one and two are the same object. The monitor is already asking your category questions on a schedule, against the competitor set the whole platform uses. The alert names a competitor you already have history on.
- Steps three and four are already gathered. What that competitor published and what people say about them are the inputs the platform watches anyway, so the cause arrives with the alert rather than after two weeks of investigation.
- Step five is a lookup, not a project. The customer groups are already computed from your own purchase data, so "does this question come from customers we want more of" is answerable the same afternoon.
- Step six starts from the brand specification. The draft is in your voice because the voice is held in the same place, so the review is an edit rather than a rewrite.
- Step seven happens because it is the same measurement. The monitor from step one keeps running. Nobody has to remember to check; the change either shows up in it or does not.
What this example is not
It is not a claim that the answer is automatic or instant. Steps five and six involve judgement that no platform supplies, and step six involves writing something genuinely worth quoting, which is the hard part and always will be. The claim is narrower: that the seven steps stop being seven separate pieces of work, and that the last one stops being optional.
Why the last step is the one that matters
A stack where steps one to six work and step seven does not is a stack that never learns. Every future version of this question starts from the same place, because the outcome of the previous attempt was never recorded against the decision that produced it.
That is the difference between a team that has been doing this for two years and a team that has done the first year twice.
Seven steps, four tool boundaries. The work does not fail at the steps. It fails at the boundaries.
If you are an agency
This walkthrough is a useful shape for scoping an engagement, whatever tools are involved. Run it with a client against their actual stack and mark which of the seven steps has a named owner. The unowned ones are almost always three, five and seven, and naming them is often worth more to the client than the work itself.
How we do this ourselves
The seven steps map onto the three parts: Signal covers one to five, Studio is six, and Proof is seven feeding back into one. You choose which parts you take, and each one is useful alone. What you get by taking more of the chain is that fewer of the boundaries need a person holding them together, and the loop closing is the thing worth testing in a demo.
The three parts, and how the loop closes