Which page of your website AI quotes, and which one it ignores

Every HVAC owner who has ever been sold marketing has been told to blog. In 2026 somebody counted which pages the AI assistants actually quote from a contractor's own site. Blog posts came last. By a lot.

Which page gets quoted, in numbers

A 2026 study published by Cheers recorded every source cited across 9,434 ChatGPT answers to home-services questions in 100 US and Canadian metro areas, between 31 August and 20 September 2026. When ChatGPT cited the contractor's own website, this is which page it took.

Page type on the contractor's site	Share of those citations
Homepage	37.4%
Service page, for example "AC repair"	24.1%
Service plus city page, for example "AC repair in Frisco"	17.4%
Location page	10.7%
Blog post	1.2%

Source: Cheers, Where Does ChatGPT Get Its Information? Local Data 2026.

Now the number that makes all of this worth your attention. Your own website is the single most cited thing you own across three of the four engines:

• Google AI Mode cited a business website in 92.5% of answers.
• Perplexity cited a business website in 89.0%.
• Gemini cited a business website in 88.3% of its answers that carried citations.
• BrightLocal's separate research puts business websites at 58% of ChatGPT's local search sources, with business mentions at 27% and directories at 15%.

So the site matters enormously. It is just that four specific page types carry 99% of the load, and the blog is not one of them. The engine by engine picture is in how the four AI assistants pick an HVAC company.

Why the homepage wins, and what it must say

The homepage gets quoted most because it is the one page an assistant is certain exists and certain is about the company. It is the fact sheet of last resort.

Which means a homepage that opens with a slogan and a stock photo of a smiling technician gives the assistant nothing to lift. Put the facts in plain text, high on the page, in real HTML:

• What you do, in the words a customer uses. "AC repair, heating repair, new system installation, maintenance plans."
• Every town and county you cover, written out. Not a map image. Not a dropdown.
• Hours, including whether you answer at night and on weekends.
• Phone number as text, not inside an image.
• License number and the state that issued it.
• Years in business, number of technicians, and the brands you service.

None of that is new advice for a human reader either. It is just that now a machine is reading it first, and the machine cannot infer what a person would guess.

The service and city page is the gap

Service pages took 24.1% of the citations and service plus city pages 17.4%. Together that is nearly as much as the homepage, and it is the part most HVAC sites simply do not have.

A typical HVAC site has a homepage, an about page, a contact page and one page called "Services" that lists eight things in bullet points. There is nothing for an assistant to quote when somebody asks about one specific job in one specific suburb.

The fix is a page per real combination you actually serve. Not a thousand generated pages, which is spam and gets treated as spam. A short list of the jobs that pay and the towns you want more work in:

• "AC repair in [town]" for your top five towns.
• "Furnace replacement in [town]" for the same towns.
• "Emergency HVAC repair in [town]" if you genuinely answer after hours.
• One page per high-ticket job: system replacement, heat pump installation, ductwork, maintenance plans.

Each page needs something true and specific on it that no template could generate. Response time in that town. A neighbourhood name. The common unit type in that area's housing stock. The permit office you deal with. That specificity is what separates a useful page from scaled content abuse, which Google's own guidance warns about.

Why your blog post gets 1.2%

Here is the honest answer, and it applies to the page you are reading too.

A blog post is rarely the best answer to "who should I call for AC repair in my city". It is a good answer to "how does AI pick a company" or "what does AC repair cost". So the engines quote blog posts for explanation questions and skip them for hiring questions. Hiring questions are the ones that pay your payroll.

A blog earns attention and trust. A service and city page earns the citation that gets you named.

That does not make a blog worthless. It makes it a different tool. The blog is where the original data lives, which is what earns links and mentions from other sites, and business mentions were 27% of ChatGPT's local sources in BrightLocal's research. But if you have ten hours and no service and city pages, do not spend those hours blogging.

What makes a page quotable

Research into generative engine optimisation gives a usable order of operations. The Princeton study presented at KDD 2024, tested against Perplexity, measured which changes raised a page's visibility in AI answers.

Change to the page	Measured visibility lift
Cite sources	+40%
Add statistics	+37%
Add quotations	+30%
Authoritative tone	+25%
Improve clarity	+20%
Keyword stuffing	minus 10%

Source: Aggarwal et al., GEO: Generative Engine Optimization, KDD 2024.

Note the last row. Stuffing your town name into a page fourteen times does not just fail. It actively lowers your visibility. Low-ranking sites gained the most from citations, up to a 115% increase in that study, which is good news for a smaller company.

Practically, for an HVAC page:

• Answer the question in the first 40 to 60 words of the section. Do not build up to it.
• Write each heading the way a customer phrases the question.
• Make every claim stand alone. A sentence that needs the paragraph above it to make sense cannot be lifted.
• Show a date. Undated pages lose to dated pages because recency is weighted.
• Name a human author with a real role.

Is your site even readable by AI?

All of the above is wasted if the crawlers cannot reach the page. Three failure modes, in order of how often they bite:

1. robots.txt blocks the AI crawlers. If GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot, Claude-SearchBot, Google-Extended, Applebot or Bingbot is disallowed, that engine cannot cite you. Many site builders add these blocks by default.
2. A firewall or bot-protection rule blocks them. Cloudflare and similar services often challenge unknown agents. The crawler sees a challenge page, not your services.
3. The facts only exist after JavaScript runs. If your service list, hours or phone number are injected by a script, a crawler may see an empty shell. Put the facts in the served HTML.

Checking all three takes under an hour and has the largest downside of anything in this post.

Do you need an llms.txt file?

You will be told you do. The honest answer in 2026 is: build it if it takes you ten minutes, and do not expect it to do anything on its own.

Ahrefs analysed 137,210 domains in May 2026 and found 97% of valid llms.txt files received zero requests that month. Of the 3% that received any request, 96% came from bots, mostly SEO audit tools. No major AI provider has publicly confirmed that its production answer system reads the file when generating an answer, and Google's John Mueller called it "purely speculative for now".

The other side of the ledger: the GEO tracking firm Profound reported that crawlers from Microsoft and OpenAI do fetch llms.txt and llms-full.txt, and Chrome's Lighthouse now includes an agentic-browsing audit that checks for it. The cost is minutes and the risk is zero.

So: yes, add it. No, do not let anyone sell it to you as the lever. The levers are the ones in the table above.

The page checklist

Run this against your homepage and your top three service pages today.

• Does the page state what you do, where, and when, in plain served HTML?
• Is the phone number text, not an image?
• Is every town written out in words?
• Is the answer to the page's own question in the first 60 words?
• Are the headings phrased as questions a customer asks?
• Is there a visible last-updated date?
• Is there a named author or owner?
• Is there one fact on the page that no competitor could copy?
• Is there a working booking link? Google AI Mode linked a booking page in 31.2% of answers.
• Are the AI crawlers allowed in robots.txt and through the firewall?

Questions owners ask

Should I stop blogging if blog posts only get 1.2% of citations?

No, but reorder the work. Blog posts were 1.2% of the citations to a contractor's own site in ChatGPT, while service pages were 24.1% and service plus city pages 17.4%. Build the service and city pages first. Keep the blog for original data that earns mentions from other sites.

How many service and city pages should an HVAC company build?

Start with your top five towns crossed with your two highest-ticket jobs, so around ten pages. Each needs something specific and true that a template could not produce. Generating hundreds of near-identical pages is treated as scaled content abuse and can hurt you.

Does my website need llms.txt to be cited by AI?

No. Ahrefs found 97% of llms.txt files across 137,210 domains received zero requests in May 2026, and no major AI provider confirms reading it at answer time. It costs minutes and breaks nothing, so add it, but the real levers are cited statistics, clear structure and crawler access.

Why does the homepage get quoted more than my service pages?

Because it is the one page an assistant is certain exists and is certain is about your company. It is the fallback fact sheet. That is why the plain facts, services, towns, hours, phone and license, need to sit in readable text high on the homepage.

Can AI read my site if it is built in a page builder?

Usually yes, but check three things: that robots.txt does not block GPTBot, PerplexityBot, ClaudeBot or Google-Extended, that your firewall is not challenging those crawlers, and that your services, hours and phone number appear in the served HTML rather than only after JavaScript runs.