What to measure when your marketing runs itself

When a person does your marketing, you measure the person: what did they ship, what happened, was it worth the retainer. When software does it, that framing breaks down quietly. Output goes up by an order of magnitude, so 'did we ship things' stops discriminating. The system acts daily, so monthly reporting hides most of what happened. And because you are supervising rather than operating, you need to know something new: whether to trust it.

This is a framework we use, split into four layers. It is opinionated and it deliberately excludes some numbers that show up on most dashboards.

Layer one: is the system healthy

Before any performance question, you need to know the machine is running and its inputs are sound. This layer is boring and it is where most silent failures live.

• Are all integrations connected and syncing? A Search Console token that expired six weeks ago means six weeks of decisions made on stale data.
• Is conversion tracking firing correctly? Optimising ad spend toward a broken tag is the single most expensive failure in small-account paid search.
• Did scheduled jobs run? Crawls, prompt checks, ads reviews.
• How many actions did the agent take, and how many were blocked by a guardrail? A rising block rate means your limits and the agent's intentions have diverged, and one of them needs revisiting.

Check this weekly. It takes two minutes and it prevents the class of problem where you discover in April that nothing has worked since February.

Layer two: leading indicators

These move within weeks and tell you whether the work is landing, well before revenue can confirm anything.

For technical SEO

Indexed page count, crawl errors trending down, average position on your tracked query set, and Core Web Vitals field data. Impressions are the single best early signal: they move before clicks do, and a rise in impressions with flat clicks is a specific, fixable problem about titles and intent rather than a general failure.

For content

Articles published against plan, time to first impression per article, and the proportion of published articles that have earned any impressions at all after sixty days. That last one is the honest quality check on a high-volume programme. If half your output never gets seen, volume is not your problem.

For AI visibility

Mention rate across your prompt set, per assistant, tracked over repeated runs. Plus citation count by page, which tells you which content is doing the work.

For ads

Cost per lead by campaign against your ceiling, search term waste as a share of spend, and impression share lost to budget. Not click-through rate on its own, which is easy to improve in ways that cost money.

Layer three: the numbers that decide budget

There are fewer of these than most dashboards suggest. For a small business the list is close to:

1. Qualified leads by source.
2. Cost per acquired customer, blended and by channel.
3. Pipeline value influenced by marketing.
4. Revenue per pound of marketing spend.

The critical word is qualified. Total leads is the metric that most reliably makes a marketing programme look successful while a business gets worse. Any system optimising toward form fills will find you more form fills, and some meaningful share will be worthless. Feed close-rate data back so the optimisation target is customers rather than submissions.

On attribution: every model is a simplification and it is better to say so than to present a weighted figure as fact. Last-click flatters whatever runs closest to the sale. First-touch flatters whatever runs earliest. We show the full path per deal alongside a weighted estimate and label the estimate as an estimate.

For most small businesses the useful question is not 'how much did this article earn' but 'does this channel appear in the paths that close, and does the pipeline thin out when we stop'. Both are answerable honestly.

Layer four: is the automation trustworthy

This layer does not exist when a human does the work, and it is the one people forget to build. You are delegating decisions; you need evidence about the quality of those decisions.

• Approval acceptance rate. If you approve ninety-five per cent of proposals, the agent has calibrated to your judgement and you can loosen supervision. If you reject a third, do not turn autonomy up.
• Reversion rate. How often does a change get rolled back, and by whom? A rising reversion rate is the clearest early warning available.
• Guardrail hit rate, as above.
• Time from detection to fix. This is where automation should be dramatically better than a human process, and if it is not, something is stuck in a queue.

Together these tell you whether to move the system from ask-first to notify-after to fully autonomous. That progression should be earned with data, not chosen on day one out of enthusiasm or refused forever out of nerves.

What to stop measuring

Some familiar numbers actively mislead once volume goes up.

• Total sessions. With more content you will get more sessions, including a long tail of low-intent visits. It goes up whether or not anything good is happening.
• Bounce rate on informational content. Someone reading an answer page and leaving satisfied is a success.
• Keyword rankings in isolation. Position four on a term nobody searches is not an achievement; impressions and clicks are the check.
• Article count as a headline. It is a capacity measure, not a performance one.
• Any single AI answer screenshot. It is one sample from a probabilistic system.

How often to look

Weekly: system health, guardrail hits, anything the agent escalated. Five minutes.

Monthly: leading indicators, approval and reversion rates, cost per lead by channel. Half an hour.

Quarterly: cost per acquired customer, pipeline influence, and whether to change the autonomy level. This is the meeting that should actually change something.

The temptation with a system producing daily data is to look daily. Resist it. Most of what you would see day to day is noise, and reacting to noise is how a working programme gets dismantled in week three.

The shift underneath all of this

Traditional marketing measurement asks whether the work got done, because the work was the constraint. When production is cheap and continuous, the constraint moves to judgement: what should be worked on, and whether the system's decisions can be trusted.

Measure those, and the output takes care of itself.