How AI Assistants Decide Which Businesses to Recommend

The question that now decides a shortlist

Ask an assistant for a good commercial electrician near Croydon who does EV chargers and you will usually get three to five names with a sentence each. The customer rings the first two. Everyone else on the high street may as well not exist for that job.

Business owners reasonably want to know how those names are picked. Nobody outside the providers knows the exact mechanics, and they change. But the systems are built from parts we do understand, and the patterns in which businesses get named are consistent enough to act on.

What happens behind the answer

For a recommendation query, most assistants do not rely on what the model remembers from training. They run a live search, often against a maps or local business index as well as the web, collect candidates, read what they find about each, and then summarise.

That means three things matter: whether you appear in the candidate set, what the model reads about you, and how you compare with the others it read about. It is closer to a customer doing twenty minutes of research very quickly than to a ranking algorithm.

Signals that consistently seem to matter

Notice that review content matters as much as the star rating. If thirty reviews say fitted our EV charger quickly and tidied up afterwards, the model has a sentence to write about you for exactly that query.

Why your competitor keeps getting named

When we look at why one firm is recommended and another is not, the reason is rarely mysterious. Typically the named firm has one or more of these:

• A dedicated page for the specific service rather than a generic services list
• Several recent reviews that use the words in the query
• A listing on the trade association's find-a-member directory
• A mention in a local roundup or comparison article that the assistant found
• Opening hours and contact details that match across every source

The businesses left out often have a better reputation offline and a thinner written footprint online. The assistant cannot recommend what it cannot read about.

A practical order of fixes

1. Audit your facts. List every place your business appears: Google Business Profile, Apple Maps, Bing Places, directories, social profiles, supplier pages. Make the core details identical.
2. Write one page per core service. Say what it is, who it is for, the areas covered, typical price range and lead time. Put the answer in the first paragraph.
3. Ask for specific reviews. After a job, ask customers to mention what you did. Do not script them or incentivise them; that breaks platform rules and reads as fake.
4. Get listed where your category is discussed. Trade bodies, accreditation schemes, relevant marketplaces.
5. Answer the comparison questions yourself. A page explaining how to choose a supplier in your field, honestly written, is often retrieved for exactly these queries.
6. Re-check quarterly. Run the same questions and note what changed.

Businesses with many branches or a large service list usually find step one the hardest to keep up, because details drift every time hours or services change. Syncing listings from one master record is a small automation job that saves a recurring headache.

The page-level side of this overlaps with generative engine optimisation, and structured data helps too; our note on structured data for online stores applies to service businesses in most respects.

What does not work

Stuffing town names into footer text, buying reviews, creating near-identical location pages for forty towns you do not really serve, and publishing fake best-of lists that happen to rank you first. Some of these used to nudge traditional rankings. They are risky now, partly because platforms remove fake reviews and partly because a model reading forty identical pages concludes very little.

Paying for a listing on a directory nobody visits also rarely helps. The test is whether real people use the site to choose suppliers in your trade.

The honest limits

Assistants vary. The same question can produce different names on different days, in different towns, and for different users. Some weight maps data heavily; some lean on web articles. You will not control this, and anyone offering guaranteed placement is overselling.

There is also a fairness issue worth knowing about: very new businesses with few reviews struggle in AI recommendations, just as they do in the map pack. That improves with time and real customers, not with tricks. When we help clients with this at SpiderHunts, most of the work is fixing information that was already wrong, which has the pleasant side effect of helping ordinary search too.

Frequently asked questions