How to Get Your Business Cited by AI

When someone asks ChatGPT "who should I hire for X in my city," the model returns a short list of businesses. Not ten blue links. A handful of names, sometimes with reasons attached. If your business is not on that list, you are invisible to a growing share of buyers.

We built an open-source tool called canonry that automates this monitoring. It asks AI models the same queries your customers would ask, then records whether they mention your business in the answer. Run it daily or weekly, and you build a dataset of how your citation visibility changes over time.

Using canonry, we tracked 11 keywords across 66 separate checks (what we call "runs") over two weeks for a local service business. Each run asks ChatGPT, Gemini, Claude, and Perplexity the same query and records whether the business gets named. The patterns are not random. Here is what actually determines whether you show up.

What the data tells us about citation volatility

One thing that surprised us early on: AI citations are not stable. A business can be cited for a query on Monday and absent on Wednesday, then back on Friday.

For branded queries (think "[business type] + [city]"), we measured citation rates between 82% and 90% across runs. That means even for queries where a site is well-positioned, the model drops it roughly 1 in 5 times. For more generic queries like "[industry] agency [city]," the citation rate drops to 31%. For purely informational queries like "how to rank on ChatGPT," it is 0%.

This is not a bug. AI models introduce randomness (called "temperature") into their responses. They also change behavior as their retrieval systems update. The practical implication: you need to be positioned strongly enough that the model cites you most of the time, not just once.

Structured data is the highest-leverage fix

We also built an open-source audit tool called @ainyc/aeo-audit that scores any website on 13 factors correlated with AI citation readiness. You give it a URL, and it checks your structured data, content structure, entity signals, and more, then returns a score out of 100. The single factor with the most weight (12 out of 100 points) is structured data.

Here is a real comparison between two sites scored with the tool:

The optimized site gets cited on 5 of 11 tracked keywords. The unoptimized site gets cited on 0 of 23.

JSON-LD schema markup gives AI models a machine-readable description of what your business is, where it operates, and what it does. At minimum, you need:

• LocalBusiness schema with name, address, phone, service area, and hours
• Service schema for each service, linked to the parent organization
• FAQPage schema on pages with Q&A content
• Person schema for founders or key team members

Google's Rich Results Test and Schema.org's validator let you check your markup before deploying. The schema markup guide goes deeper with copy-pasteable examples for each type.

Content extractability matters more than content length

A surprise from the audit data: content depth (word count, heading structure) matters less than content extractability (how easy it is for a model to pull clean facts from your page).

The optimized site in the comparison above scores 87 on content depth but only 65 on extractability. The unoptimized site scores 72 on depth but 45 on extractability. Both have decent amounts of content. The difference is how that content is structured.

What makes content extractable:

• Definition blocks. Start key pages with a clear "X is Y" statement. If someone asks "what is [your service]," the model needs a sentence it can pull directly. The unoptimized site in the comparison scores 0/100 on definition blocks because none of its pages open with a direct definition.
• Question-based headings. Use H2s that match how people ask questions. "How much does roof coating cost?" maps directly to how models parse content for answers.
• Short paragraphs. Two to four sentences each. Models extract paragraph-level chunks. Walls of text are harder to parse.
• Lists and tables. Models extract structured formats more reliably than prose.

Sites built with heavy page builders (Elementor, Divi) often score poorly on extractability because content is buried under layers of wrapper divs. The aeo-audit tool's Content Extractability factor measures content-to-markup ratio specifically for this reason.

Entity consistency is the silent killer

Entity consistency scored 86/100 on our optimized site and 42/100 on the unoptimized one. This is the factor that most businesses overlook because it is not visible on their own website.

AI models cross-reference your business across the web. If your business name, address, phone number, and service descriptions are inconsistent across your website, Google Business Profile, Yelp, and directories, the model has lower confidence in recommending you.

This is the same NAP (Name, Address, Phone) consistency that local SEO has emphasized for years, but it matters even more for AI because models use entity resolution to decide whether multiple web mentions refer to the same business.

Concrete steps:

• Audit your listings on Google Business Profile, Yelp, industry directories, and social platforms
• Make sure the business name is exactly the same everywhere
• Use the same phone number format consistently
• Link your website to all major profiles

Semrush's listing management tool and BrightLocal both help with auditing consistency at scale.

Publish an llms.txt file

llms.txt is an emerging standard that tells AI crawlers what your site is about and where to find key information. The optimized site in the comparison scores 100/100 on AI-readable content partly because it has both llms.txt and llms-full.txt.

A minimal llms.txt includes your business name, what you do, where you operate, and links to your most important pages. This is low effort and high signal. Some WordPress SEO plugins auto-generate a basic version, though you will want to customize it.

Get indexed first

The exact retrieval systems behind each AI model are not fully public, and they change frequently. Based on what we have observed and what has been announced: Gemini appears to pull from Google's index for grounding. ChatGPT has used both Bing and Google for web browsing. Claude has its own web search capability. Perplexity runs its own real-time search. The safest approach is to be indexed everywhere. Submit your sitemap to both Google Search Console and Bing Webmaster Tools.

Check your index status in Google Search Console. Use the URL Inspection tool to request indexing for new pages. Canonry also integrates with GSC, so you can check indexing status from the same tool you use for citation monitoring.

Monitor, do not guess

The gap between "I think I am showing up" and "I am actually showing up" is where most businesses waste time optimizing the wrong things.

This is what canonry is built for. It runs queries against multiple AI providers on a schedule, tracks citation state over time, and identifies when you gain or lose visibility. The loss/recovery patterns described above only became visible because the tool was running daily sweeps automatically.

For a point-in-time assessment, the @ainyc/aeo-audit tool scores your site across all 13 factors in under 30 seconds:

npx @ainyc/aeo-audit@latest "https://yourbusiness.com" --format json

Both tools are open source. The monitoring gap in AI search is real, and we would rather help businesses close it than sell them something they could build themselves.

The realistic timeline

AI models do not update in real time. When you make changes to your site, crawlers need to re-visit, indexes need to update, and models need to incorporate the new data. This could be weeks or months.

Based on what we have observed in canonry monitoring data, the typical pattern is:

1. Week 1-2: Changes deployed, indexing requested
2. Week 2-4: Pages start appearing in search indexes
3. Week 4-8: Citation patterns begin shifting in AI answers
4. Ongoing: Citation rates stabilize but continue to fluctuate (remember the 82-90% rate, not 100%)

The right approach is positioning and monitoring. Make your site the best possible candidate for citation. Then track what happens. Canonry will tell you if and when it pays off.