How we built a local business site that AI engines cite

Why do AI engines ignore most local business sites?

Because most local sites give them nothing to work with. An AI engine deciding who to recommend needs three things: access (can its crawler fetch the pages?), structure (can it parse what the business does and where?), and corroboration (do the facts match what Google, Bing, and the review platforms say?). A template site with five thin pages fails all three quietly, and the engine names someone else.

100th Meridian started there, on a small Wix site. That gap costs more every year: roughly 1 in 5 people now start searches on an AI platform, according to Quantum Metric's 2026 benchmark report. The finished site is live at 100thmeridianwindowfilm.com, and the results are in the case study.

Decision one: architecture before content

The rebuild is a static site with one page per service and one per service-area city, linked within each group. The services are heat-control film, privacy, security, decorative and UV. That replaces one homepage that had to rank for everything.

Why it matters to AI specifically: research on what generative engines absorb found that top-cited pages average 1,943 words with 12.5 times more headings than bottom-quartile pages, and that pages carrying statistics show 61.55% higher influence on answers (Zhang et al., arXiv, 2026). Depth and structure are the raw material of citations. A silo gives every buyer question a substantive page to cite instead of a paragraph buried on a homepage.

On launch day every page was submitted for indexing, so the engines did not have to find the new structure on their own. How quickly Google and Bing indexed it is recorded, with dates, in the case study.

Decision two: make every fact machine-readable

Every page ships structured data (LocalBusiness, Service and FAQ schema), a clean heading hierarchy and accurate metadata. The business's name, phone, service list, and area are stated identically on the site and its Google Business Profile, because engines cross-check before they recommend. When facts disagree, the safe move for the model is to name a competitor whose facts agree.

This is the least glamorous work in the stack and the one everything else rests on. Reviews are part of it: a study Trustpilot commissioned from Seer Interactive (May 2026) analyzed more than 800,000 AI answers: brands that collected more than 80 Trustpilot reviews and replied regularly were cited in 75.3% of AI answers, against 1% for brands with no active Trustpilot profile. Trustpilot sells review profiles, so read it as the vendor's own research. It is standard technical SEO applied with AI retrieval in mind: schema so machines parse rather than guess, the same name, address and phone on every listing. We also ship an llms.txt file, but few engines read it, so it rides along as a cheap extra rather than a pillar.

Decision three: verify who is actually reading the site

From day one the site runs our bot tracker, which checks each claimed AI crawler against its operator's published IP ranges. It's the same approach Google documents for verifying its own crawlers. Raw server logs overcount: scrapers send requests that claim to be GPTBot or ClaudeBot, and on this site a large share of the requests claiming to be AI or search bots failed that check. Count them and you report AI interest that never happened.

Verified traffic shows which engines are really reading the site and how soon after launch they arrive. The one to watch is ChatGPT-User, which OpenAI describes as the agent for user-initiated actions, not automatic crawling. A verified ChatGPT-User visit means a person's request in ChatGPT led it to fetch the client's page. The verified counts for this site, crawler by crawler, are in the case study.

Decision four: measure AI visibility like a metric, not a vibe

During the launch period the runs were 3 to 7 days apart; today we run it every two weeks. Each run asks ChatGPT, Claude, Gemini, Perplexity and Google AI Mode the questions a window film buyer in the client's market would ask, about 20 per chat engine plus local keywords for AI Mode, and log every business each answer cites. A citation only counts when the engine links the client's own site. Share of voice is the client's share of the AI citations that went to any named business in the market, so soft questions and vague mentions cannot flatter it.

That harness turns a launch from an anecdote into a trend you can check run by run. The run-by-run figures for this client, each with its date, are in the case study.

Decision five: count only leads that provably happened

The site tracks a lead only when the server confirms an inquiry email was actually delivered. Phone taps and email clicks are tracked separately as weaker evidence of a lead, and every QA visit we make carries a test flag that excludes it from the stats. A phantom lead is worse than no lead, because it teaches the client to trust numbers that are not real.

Rules that strict keep the counts small, and every one of them is a real inquiry. The inquiry and phone-tap figures for this site, with their dates, are in the case study.

What transfers to your business, and what does not?

What transfers is the input list, because none of it is specific to window film. Real pages for every service and city, machine-readable facts that agree everywhere, verified measurement, and a launch process that submits everything for indexing on day one: that is the same playbook we follow inside our program, whether we are fixing an existing site or rebuilding one, which is an add-on quoted separately.

What does not transfer is the slope. This is one company in one market, so read it as evidence that the method can work, not as a forecast. Niagara window film turned out to be a winnable field, and a crowded market moves slower. Why AI answers favour businesses whose facts are clear and consistent is covered in how to show up in ChatGPT results.

If you want to know where you stand, the free AI visibility check shows whether AI crawlers can reach your site and read its business schema, and asking ChatGPT "best [your service] in [your town]" costs nothing at all. If someone else's name comes back, the five decisions above are the list of what to fix.

Frequently asked questions

How long does it take for AI engines to cite a new local business site? There is no fixed timeline, and one site is not a forecast. On this build, verified AI crawlers arrived soon after the July 11, 2026 launch, and the case study dates each result that followed. A site the engines cannot parse gives them nothing to cite, however long you wait.

Do you need a full rebuild to show up in AI answers? No. The same layers (dedicated service and city pages, schema, consistent business facts, verified measurement) can be added to an existing site. A rebuild, quoted separately as an add-on, is the right call only when the platform blocks that work, and the Diagnosis is where we find that out.

What is a verified AI crawl? A request whose claimed crawler identity (like GPTBot or ClaudeBot) is confirmed against the operator's published IP ranges. Counting unverified hits inflates AI visibility with scraper traffic.

How is AI share of voice measured? By asking ChatGPT, Claude, Gemini, Perplexity and Google AI Mode the same set of buyer questions on a fixed schedule (3 to 7 days apart during the launch period, every two weeks today) and logging every business each answer cites with a link. Share of voice is one business's share of the AI citations that went to any named business in the market, so soft questions cannot inflate it.