Scoring local businesses with AI: what to tell it

Tell the AI four things: what your best customer looks like in facts it can see, what rules a business out, what to do when a fact is missing, and what a high score and a low score mean. Give it those facts by putting the right blocks before it, ask for a score from 0 to 100 with a one-sentence reason, and read the reasons on a test run before you trust the numbers.

This is for anyone who has a list of local businesses and more of them than there are hours to call: distributors, agencies, software sellers, suppliers. It explains what the AI actually reads, how to write instructions that produce useful scores, how to check them, and when a different kind of question works better than a score.

What the AI can and cannot see

The xMap LLM block reads, for each business, the fields the blocks before it found: name, category, address, rating and number of reviews, whether it is a chain, phone and website, the date it opened when known, and anything else earlier blocks added, such as how many rivals are nearby, how many people live around it, or the job a company is hiring for. It also reads short notes you write and the text of reference files you attach: PDF, TXT, CSV, Markdown or JSON, up to 3 MB each, with about 40,000 characters of text read from up to 10 files.

It does not browse the web or look a business up elsewhere. If your scoring depends on something, that something has to be in the flow. Want it to judge how busy the street is? Put a block before it that counts the places around each business. Want it to know your price bands? Put them in the notes. Long texts are cut to 300 characters to keep reading fast, with reviews and job descriptions the exceptions, read much longer.

Instructions that work

Here is a made-up example: a coffee roaster in Melbourne that sells beans wholesale to independent cafes. A weak instruction looks like this:

Score these cafes as leads.

The AI will do something, but it has to guess what a good lead is, so it will lean on the star rating. A strong instruction says what the roaster's salespeople already know:

We sell specialty coffee beans wholesale to independent cafes. A great prospect (80 to 100) is an independent cafe, not part of a chain, with a rating of 4.3 or more from at least 50 reviews, ideally opened in the last two years, since newer cafes are still choosing a roaster. A fair prospect (50 to 79) is an independent with a good rating but few reviews, or an older cafe with a strong rating. Score under 30 any chain branch, any place that is mainly a restaurant or bar, and any cafe already selling a named roaster's coffee in its name. If a fact is missing, say so in the reason and do not guess. Give one short sentence of reason a salesperson can read aloud.

That instruction works because every condition is something in the data: the chain flag, the rating, the reviews, the opening date, the category, the name. It sets the scale with words and numbers, and it says what to do about gaps, which stops the AI from inventing a reason.

Notes and reference files

Keep the instruction about the judgement, and put the background in the notes and files. Good notes are short facts: the areas your vans deliver to, the minimum order you accept, the kinds of business you never sell to. Good reference files are things like your product list, so the AI can match products to a place, or a description of your three best customers. A file of a thousand customer names is less useful than a paragraph describing what they have in common.

Check the scores before you trust them

In your account, every run uses the AI for real, test runs included, so you can tune without sending anything. Run a test, open the results, and read ten reasons from the top, ten from the middle and ten from the bottom.

• If the top reasons are all about star ratings, your instruction is not telling the AI what else matters.
• If many reasons say a fact is unknown, the flow is missing a block that would provide it.
• If the scores bunch between 60 and 75, tell the AI more firmly what a 90 and a 30 look like.
• If one result looks wrong, use Ask why on it: the answer comes from that result's own fields, its score and reason, and the rules the flow applied.

Change the instruction, test again, and compare. Two or three rounds are usually enough.

From scores to a call list

A score only helps if it changes what happens next. After the AI step, use Filter with Fit score to keep, say, 70 and above, and Sort and limit to put the highest first and keep as many as a rep can call in a day. Put Only new before the AI step, not after, so the flow only scores businesses it has not seen before. In the results, any field with score, fit or rank in its name is read as a number from 0 to 100, and the run's summary line counts how many strong fits it found.

When a score is the wrong question

Sometimes you do not need a ranking but a decision. Two other ways work better then:

• The Classify task in xMap LLM gives each business one of your own labels, such as Call this week, Email, or Not for us, with a reason.
• The Jev decision block asks one typed question of each business: yes or no, one of your options, or a score on a few levels, with a confidence.

Mistakes that spoil a score

• Asking for facts the flow does not have, such as revenue or number of staff, then trusting the guess.
• A scale with no anchors. Without examples of a 90 and a 30, scores drift to the middle.
• Too many fields at once. Ask for a score and a reason first; add more fields only when you use them.
• Treating the score as a fact. It is a judgement on the data in front of it. Reps should still look before they call.