Read your first results

The first run arrives minutes after the first schedule fires. It is not a baseline. This page is about what to read on day one, what to leave until day six, and what to do first.

Day one: read the answers, not the charts

Open Runs, pick any run and read the full answer the model produced. Do this ten times. It is the highest-value half hour of the whole setup, because it tells you three things no chart can.

• Whether your prompts sound like real questions. Models hedge on vague ones, and a hedged answer names no brands at all.
• Whether your brand and your competitors are being recognised, which you then fix in Name matching.
• What assistants currently believe about your market, in their own words. This is usually the part you end up quoting to other people.

Day one: read the fan-out searches

Open AI searches. These are the searches each model ran on its own while answering your prompt, and they are what decided which sources it read.

They are also the most actionable screen in week one, because they need no trend data. A search that appears across most of your prompts, with no page of yours answering it, is a page to write this week.

Day six: read the overview

By now each prompt has several batches and the numbers mean something.

• Visibility Score, the headline: the share of grounded probes where the model recommended you, with its confidence band.
• Visibility Score by Platform: the same thing per assistant. The spread between assistants is usually wider than any week-on-week move.
• Share of voice: your share of the brands named back, against the competitors you registered.

Grounded against training

Every figure can be read in three modes, and the Visibility Score uses the first.

• Grounded: the prompt ran with web search available, so the answer reflects what the model can find today.
• Training: it ran with no search, so the answer reflects what the model already believes about you.
• All: both together.

A brand that scores well grounded and badly on training is winning on retrieval and has not yet made it into what the models know by default. The reverse means your reputation is ahead of your pages. Those are different problems, and grounded and training covers both.

The first thing worth doing

In order, and stop at the first one that applies.

1. Names mapped wrong. Fix them in Name matching. Everything downstream is wrong until this is right.
2. Prompts that read like keywords. Rewrite them. Bad prompts produce clean charts of nothing.
3. No mentions anywhere. Start on retrieval readiness: the question is whether your pages can be found and quoted at all.
4. Named on some assistants, invisible on another. Compare the cited sources on each. The gap is almost always a source the invisible one leans on that does not carry you.
5. A healthy grounded score and nothing on training. Keep going. That gap closes over months, not weeks, and it closes through being cited repeatedly.

What good looks like after a month

A Visibility Score you can state with a number and a band, a trend that moves when you ship and sits still when you do not, and a control prompt proving the difference. That takes a month rather than a week.