Google's generative AI performance report counts how often links to a site appear in AI Overviews and AI Mode — but it cannot tell a business whether the answer recommended it or a competitor. That gap sits at the center of how local businesses now get chosen.

When a customer asks an AI assistant for a local recommendation, the assistant returns a short list it has already selected, before the customer reaches anything the business controls. Google positions AI Mode as suited to questions requiring exploration, reasoning, or complex comparisons — the kind it says previously took several searches.

In organic search, a business could check rankings and know roughly where it stood. AI answers compress that. Uberall's quick-service restaurant benchmark typically produced three to five recommended brands per query, according to the vendor's own reading of one industry. That indicates how small the field of visible options has become.

What Google Confirms About How Answers Are Built

Google says its AI features in Search draw on the same ranking and quality systems as regular results, and it names two techniques specific to answer generation.

The first is retrieval-augmented generation (RAG), which Google also calls grounding. It uses existing ranking systems to pull relevant pages from the Search index; the AI generates its answer from what those pages say and links to the supporting pages.

The second is query fan-out, a technique that runs multiple related searches at the same time and gathers more results than the original question would return alone.

For a business's content to contribute, it must sit on a page Google can retrieve. If someone asks for a quiet restaurant for a client lunch, that quality has to be described somewhere Google can read — in the reviews or the business description. When Google pulls multiple sources with conflicting information about the same business, it does not document how it resolves the difference, so consistent information across every controlled page reduces the risk.

Listings accuracy, an active review profile, and consistency across sources all affect whether an assistant considers a business at all. Meeting Google's requirements, however, does not guarantee inclusion in AI-driven features.

The Measurement Gap, Quantified

Google folds AI Overviews and AI Mode data into the main Search Console Performance report under the Web search type, mixed with everything else; those clicks cannot be separated there. The generative AI performance report does separate them, and Google is still rolling it out — not every property has it yet. It counts impressions and breaks them down by page, country, device, and date.

What it lacks is a query dimension. An impression says a link appeared, nothing more. It does not distinguish an answer that recommended the business from one that recommended a competitor and cited the business as a supporting source. The report also covers Google only — nothing in Search Console shows what ChatGPT, Perplexity, or Claude told someone asking which local business to use.

Third-party data fills part of the gap, with caveats. Whitespark's analysis, published in July, tested 540 queries across three U.S. cities and six industries. It found AI Overviews on 15% of direct local-intent queries, 92% of informational ones, and 97% of hybrid queries — questions like "should I hire a lawyer after an accident" that carry a purchase decision inside an information request. Those hybrid queries are where an AI answer can shape which businesses a customer considers.

Teams are gauging performance two ways while reporting catches up. Some run a fixed set of local queries against the major assistants on a schedule and record which businesses get named; because assistant answers change with phrasing, location, and session, those numbers are directional. Others separate AI referral traffic in analytics and track volume and behavior over time — which shows what happened after a click, a smaller group than everyone who saw the answer.

Google's guidance also warns against third-party tools claiming access to its internal ranking or AI systems.

Making Content Eligible

Google's AI optimization guide sets the entry requirements. A site must be included in Search generative AI features to be eligible to appear in them. Google is rolling the include/exclude setting out to some site owners and not others; "include" is the default, and a child property inherits from the closest parent with a manually configured setting unless the child overrides it. Businesses running each location as its own property should check those too, since a location property can inherit an exclusion from the domain above it.

The guide recommends keeping crawling permissions open in robots.txt and at the CDN, making important content available as text, using structured data aligned with visible page content, and keeping Business Profile details current. Google can process JavaScript when it isn't blocked, but its guide notes JavaScript-framework sites are generally more complex to work with — each rendering step is another point where a business's hours could fail to appear.

AI responses can also pull from pages the business doesn't control: an outdated directory with last year's hours, an old article, or a competitor's comparison page describing the business in the competitor's terms.

Google's guide also corrects misconceptions: Search ignores llms.txt and similar AI text files, and there is no need to break content into small chunks for Google's AI systems to understand it.

Agents Are Next

Google describes AI agents as systems that carry out tasks for people, such as booking a reservation or comparing product options. It says browser agents may access a site directly, analyzing rendered screenshots, inspecting the DOM, and interpreting the accessibility tree that screen readers rely on. An unlabeled div button, a form field identified only by placeholder text, or a phone number trapped inside an image are accessibility problems — and potentially agent problems. Google hasn't said such defects will stop an agent, but the accessibility tree is one of the layers it says browser agents may interpret.

Google points to a web.dev guide for preparing sites for current browser agents and describes protocols like the Universal Commerce Protocol (UCP) as emerging. For a local business, readiness likely means the same clear information, plus a booking or ordering system that works without anyone fixing a broken form.

Where That Leaves Local Businesses

Customers now see a curated short list before anything the business controls. Google's performance report is still rolling out, tracking tools are nascent, and neither Google's public guidance nor Search Console explains why one local business gets named over another. Until that changes, the priorities are fixed: accurate, consistent information across sources, published as text on pages Google can crawl.

Watch for the generative AI performance report reaching more properties, and for any Google documentation of how conflicting sources are resolved — the two changes most likely to make AI visibility measurable rather than inferred.