How to measure whether your website assistant helps

Decide what "helps" means before you open a dashboard

An assistant on your website can help in a few different ways, and they are not measured the same way. It might mean fewer repetitive emails. It might mean people find the right page without asking. It might mean someone gets an answer at eleven at night instead of giving up.

Pick the one that matters to you and write it down. Otherwise you will end up reading whatever number happens to be largest.

Be clear-eyed about the limit, too: Linkd sees conversations on your site. It does not see your revenue, your sales pipeline, or what someone did after they left. Anyone telling you a chat tool measured their return is measuring something else and attributing it.

What the analytics actually show

Conversations and questions. How many questions came in over 24 hours, 7 days or 30 days, and how that compares with the previous period. A monthly meter shows questions used against your plan's allowance and projects where the month will land.

Where questions come from. A split between Visitors and Testing, so your own playground runs are never counted as customer demand. Check this first. A surge that turns out to be you is a common and embarrassing mistake.

Answered from sources. The share of visitor questions the assistant could ground in your content. This is a coverage measure, not a correctness measure — it tells you the answer came from something you connected, not that the something was right or current. Use it to find gaps, and keep testing for accuracy by hand.

Satisfaction. The proportion of thumbs-up among the people who rated an answer. Treat it as a weak signal: almost nobody rates, and the ones who do are disproportionately delighted or annoyed. Read it alongside the number of ratings it is based on, which is shown next to it.

Recent questions. The questions themselves, filterable by Visitors, Testing and Unanswered. This is the most useful screen in the product and the least numeric.

Knowledge gaps. Questions that found no matching source, with how often each was asked and a direct link to write the answer as a Q&A pair.

Chat logs are off until you turn them on

Question text is only stored if you switch chat logs on, which means the counts above work from the start but the words don't. Turning logs on is what makes "recent questions" and "knowledge gaps" useful.

That is a real decision, not a formality. You are choosing to retain what visitors typed, which may include things they volunteered that you would rather not hold. Decide deliberately, tell people in your privacy notice, and keep the retention period as short as it can usefully be.

A monthly routine that takes about twenty minutes

1. Check the visitor/testing split. Make sure you are looking at customers.
2. Read thirty recent questions. Not the summary — the questions. You will recognise your business in a way no chart conveys.
3. Work the unanswered filter. Every repeated question with no source is a decision: write the answer with Teach answer, connect the page, or decide it belongs with a person.
4. Clear the top three knowledge gaps. Write them as Q&A pairs. These are the highest-value twenty minutes in the whole routine, because they are questions real customers asked that you currently cannot answer.
5. Spot-check five answers against your current facts. Prices and hours drift; indexed content doesn't notice.
6. Look at the questions allowance. If you are on pace to run out, decide now rather than on the 28th.

Pair it with a number Linkd can't see

The measure that usually matters most lives in your own systems. If the goal was fewer repetitive emails, count them for a fortnight before launch and a fortnight a month after. If the goal was fewer calls asking about hours, ask whoever answers the phone.

It won't be a controlled experiment and you shouldn't present it as one. Other things change at the same time. But a rough before-and-after you measured yourself is worth more than a precise percentage someone else invented.

Signs it isn't working

• Visitors ask one or two questions and leave without opening anything.
• The same question keeps appearing under Unanswered a month after you first saw it.
• Answered-from-sources is high but your spot-checks keep finding stale facts.
• Nobody is asking anything, which usually means the bar is hidden on pages nobody visits, or switched off, rather than that nobody has questions.

Each of those has a fix, and none of the fixes are in the dashboard. They are in your content.

Start with preparing your content.