Three Questions to Ask Your Analytics AI Today
The hardest part of an AI analytics assistant is the blank text box. Here are three questions worth typing into it right now — what each one makes the assistant do, and how to tell whether the answer came from your data or from thin air.
Key Takeaways
- Start with a period comparison: it exercises the trend logic and shows you the tool calls behind the answer.
- Then a drill-down with reasoning — adding "and why do you say that?" forces the assistant to expose its evidence.
- Then an engagement question. Bounce rate in SealMetrics is entrances minus engaged entrances over entrances; an engaged session has more than one pageview.
- Always ask for the why. It is the fastest way to verify the numbers came from your data rather than from the model's guesswork.
Seal AI answers questions about your analytics in plain language. There is no query builder to learn and nothing to configure — but there is a skill to it, and the skill is mostly about asking precisely.
These three questions are a good first session. Each one exercises a different part of the assistant, and each one gives you a way to check the answer.
1. The period comparison
"Compare my entrances over the last 7 days to the previous 7."
What it does under the hood. The assistant resolves both windows against your site's timezone, then calls the overview tools twice — once for each period — and computes the delta from what comes back. It does not remember last week's number; it fetches it.
How to read the answer. Check that the two windows are the ones you meant. Then look at the direction and the size of the change together: a small percentage swing on a low-traffic week is noise, and a good answer will say so rather than dramatise it. If the assistant offers a chart alongside the text, the chart and the prose should agree — if they do not, ask again.
This is also the cheapest way to see the mechanics. A period comparison forces at least two separate data retrievals, so it is a clean demonstration that the assistant works by looking things up.
2. The drill-down that asks for reasoning
"Which of my top 3 sources converts best, and why do you say that?"
What it does under the hood. This is a multi-step question. The assistant has to rank your traffic sources, pull conversions for each of them over the same period, divide one by the other, and then order the result. That is several tool calls chained together, with the output of one feeding the input of the next.
Why the "why" matters. Those four extra words change what you get back. Without them you receive a conclusion. With them you receive a conclusion plus the figures it rests on — which source, how many entrances, how many conversions, over which window. That is the material you need to sanity-check the claim, and it is the single most useful habit to build when working with an analytics assistant.
How to read the answer. Look at volumes before rates. A source with a spectacular conversion rate on a handful of visits is a curiosity, not a strategy. A good answer surfaces both; if it only gives you percentages, ask for the absolute numbers.
3. The engagement question
"What's my bounce rate this month and how does it compare?"
What it does under the hood. The assistant retrieves entrances and engaged entrances for the current month and for the comparison window, and derives the rate from them.
The definition you need to know. In SealMetrics, bounce rate is:
(entrances − engaged entrances) ÷ entrances × 100
An engaged session is one with more than one pageview. Two consequences follow. First, bounce rate here is the inverse view of engagement, not the same number as engagement rate — they answer different questions and should not be used interchangeably. Second, because analytics platforms define bouncing differently, this figure is not comparable with a bounce rate from another tool. Compare it against your own history, and it will tell you something real.
How to read the answer. Movement matters more than level. A high bounce rate on a page whose whole job is to answer one question is fine. The same rate appearing suddenly on a category page is a signal.
How to ask better questions
- Name the period explicitly. "Last month" is ambiguous — it can mean the previous calendar month or the last 30 days. Say which one. We know this one from experience: a benchmark question of ours phrased exactly that way accounted for every fact miss in an entire test run.
- Ask for the reasoning in the same breath. Append "and why do you say that?" or "show me the numbers behind it".
- Ask for the breakdown you want. By channel, by device, by landing page, by campaign. Naming the dimension removes a guess.
- Ask one thing at a time. Three questions in one message usually produce one good answer and two thin ones.
- Follow up rather than restart. "Now split that by device" is a better second message than retyping the whole question.
One caution
Always ask for the why. Not because the assistant is untrustworthy, but because the evidence is the part you can check and the prose is not. Every figure the assistant states should be traceable to a metric, a period and a breakdown it just retrieved from your data. When it can show you that chain, you can act on the answer. When it cannot, you have an opinion in a confident voice.
For a one-off test of the same principle, ask about a campaign you never ran. The right answer is that there is no data for it — which is exactly the behaviour we test for with grounding traps.
Where to try it
The assistant is already in your SealMetrics account — there is no key to paste and no AI vendor account to create, and inference runs in the EU only with nothing retained by default. Open a site, open the assistant, and start with the first question above.
If you want the detail underneath it first, the Seal AI architecture documentation covers how it processes your data, and Meet Seal AI is the short introduction.