Local search used to leave the comparison work to the customer. Someone typed a query, scanned a list of ten blue links, opened a few, and picked one. Ask an AI assistant for a local recommendation today and you get something different: AI-generated local business recommendations that have already been narrowed down for you, often before you see anything the business itself controls.

Google frames this as a feature, describing AI Mode as especially useful for questions that need exploration, reasoning, or comparison — the kind of thing that previously took several separate searches to work through. For local businesses, that framing matters less than the practical consequence: an assistant is now doing the shortlisting, and the criteria it uses aren’t fully visible from the outside.

Being listed correctly is the entry fee, not the strategy

Listings, reviews, and reputation still determine whether a business is even in the running before an assistant starts comparing. That groundwork hasn’t changed — accurate hours, consistent business details, and an active review profile remain the baseline, as covered in reviews, reputation, and listings as the local signals AI now reads. Meeting those requirements makes a business eligible to be considered. It doesn’t guarantee it gets recommended.

How Google says it builds the answer

Google says its AI features in Search sit on top of the same ranking and quality systems used for regular results, and it names two specific techniques behind how answers get assembled. Retrieval-augmented generation, which Google also calls grounding, pulls relevant pages from the Search index using existing ranking systems, then generates an answer from what those pages actually say — linking back to the sources it drew from. Query fan-out runs several related searches simultaneously, gathering more material than the original question alone would surface. Readers who want the mechanics behind grounding can find a deeper breakdown in this walkthrough of grounding and retrieval-augmented generation.

The practical implication: for a business’s own content to shape an AI answer, it has to live on a page Google can actually retrieve. If someone asks for “a quiet restaurant for a client lunch,” that detail needs to exist somewhere readable — in a review, a description, anywhere Google’s systems can pull it from. And because Google doesn’t document how it resolves conflicting details about the same business across different sources, keeping information identical everywhere is the safest hedge. One vendor’s benchmark of quick-service restaurant queries found the AI typically surfaced only three to five recommended brands per query — a useful reminder of how narrow the shortlist actually is.

What Search Console shows, and what it can’t

Google mixes AI Overviews and AI Mode clicks into the standard Performance report under “Web search,” so those clicks aren’t broken out there. The separate Generative AI performance report does isolate them — counting impressions, meaning how often a link to a page showed up in an AI feature — broken down by page, country, device, and date. It’s still being rolled out, so not every property has access yet.

What it doesn’t include is a query dimension. An impression tells you a link appeared; it doesn’t say whether the AI recommended that business or listed it as a secondary source underneath a competitor. And the report only covers Google — nothing in Search Console reflects what ChatGPT, Perplexity, or Claude tell someone asking which local business to use.

That gap matters more than it might seem. A separate analysis of 540 queries across three U.S. cities and six industries found AI Overviews appearing on just 15% of direct local-intent searches, but 92% of informational ones and 97% of hybrid queries — questions like “should I hire a lawyer after an accident,” which carry a purchase decision hidden inside an information request. Those hybrid queries are exactly where an AI answer is most likely to shape which businesses a customer even considers, and they’re easy to underweight if a business is only watching local-intent keywords. See why AI answers about your locations are often wrong for what happens when those answers pull from outdated or inaccurate sources.

Until better reporting exists, most teams are triangulating two ways: running a fixed set of local queries against major assistants on a schedule and logging which businesses get named (directional, since answers shift with phrasing, location, and session), and separating AI referral traffic in analytics to track what happens after someone actually clicks through — which only captures a fraction of the people who saw the answer.

Give the assistant something to quote

Google’s guidance for getting included in AI features starts with a basic eligibility check: a site has to opt into Search’s generative AI features to be eligible at all. “Include” is the default, but the setting is rolling out unevenly, and child properties inherit whatever setting the closest parent has — meaning a single location page can inherit an exclusion set at the domain level without anyone noticing. It’s worth confirming multi-location businesses haven’t inadvertently excluded individual properties, a topic covered in more depth in the complete guide to local SEO for multiple locations.

Beyond eligibility, Google recommends keeping crawling open in robots.txt and at the CDN level, making important content available as plain text, and keeping structured data aligned with what’s actually visible on the page. Google can process JavaScript-rendered content when it isn’t blocked, but its own guidance calls JavaScript-heavy sites more complex to work with — every rendering step is another place hours or services can fail to appear correctly. And Google’s guide explicitly debunks two common assumptions: it ignores llms.txt files, and there’s no benefit to artificially chunking content for AI systems to parse it.

When agents start doing the booking

Google describes AI agents as systems that complete tasks on a person’s behalf — booking a table, comparing options, placing an order — and says browser agents may read a site by analyzing rendered screenshots, inspecting the DOM, or interpreting the accessibility tree that screen readers rely on. That last point raises the stakes on basic accessibility work: an unlabeled button built as a div, a form field identified only by placeholder text, or a phone number that exists solely inside an image are exactly the defects that break for screen readers — and potentially for agents reading the same tree. For a local business, being ready for this means the same accurate, plain-text information described above, plus a booking or ordering flow that works end to end without someone needing to manually rescue it.

What this means in practice

Customers asking an assistant for a recommendation now see a short, pre-filtered list before they reach anything the business controls directly. Reporting tools for tracking this are still early, and neither Google’s public guidance nor Search Console explains why one business gets named over another in a given answer. Until that changes, the more durable priorities are the same ones local SEO has always emphasized — accurate, consistent information available as text on pages that can be crawled — applied with more discipline than before, because the room for being second-guessed by an assistant reading a stale directory listing is real.

Frequently asked questions

How do AI assistants generate local business recommendations?

Google says its AI features in Search use the same underlying ranking and quality systems as regular results, layered with retrieval-augmented generation to pull from indexed pages and query fan-out to gather more results than a single search would return. Being listed accurately and consistently is what makes a business eligible to be considered — it doesn’t guarantee it gets named.

Does Google Search Console show which AI assistant recommended my business?

Only partially, and only for Google. The Generative AI performance report shows link impressions by page, country, device, and date, but it has no query dimension, so it can’t distinguish a recommendation from a passing mention. It also doesn’t cover ChatGPT, Perplexity, Claude, or any assistant other than Google’s own.

Do I need an llms.txt file for my business to show up in AI answers?

No. Google’s own optimization guidance says Search ignores llms.txt and similar AI text files entirely, and there’s no benefit to breaking content into small chunks for its AI systems to read.

How can a local business track its AI visibility today?

Most teams combine two approaches: running a fixed set of local queries against major assistants on a schedule and logging which businesses get named, and separating AI referral traffic in analytics to see what happens after someone clicks through. Both are partial — assistant answers vary by phrasing and session, and referral tracking only captures people who actually clicked.