Your website looks great. ChatGPT can’t read a word of it.
A lot of the propane websites built in the last two years are JavaScript apps. A person sees a finished page. A crawler gets a mostly empty file and a note saying the content is still loading. Google will eventually run that code. Vercel and MERJ counted more than 500 million GPTBot fetches across Vercel’s network in a single month, and in the same study found that GPTBot — the crawler behind ChatGPT — doesn’t run JavaScript at all.
This page is every check we run on a propane website, what’s actually true about each one, and what PIMS does about it.
Company, website, ZIP. About sixty seconds. Add your email and we’ll send you the report. Nothing to buy at the end of it.
Or call us: (610) 228-0887
The ten minutes
You spent ten minutes with an AI and got a website. Clean layout. Good photos. Looks right on a phone. Three weeks later, you search your own county and you’re nowhere.
So somebody tells you AI content doesn’t rank.
That’s not it. AI content ranks fine. The problem is that nobody told the AI what a machine needs in order to crawl your site, understand it, and trust it. It built the part you can see and skipped every part you can’t.
There are twelve of those parts. The first one means an AI assistant may never have read your site at all.
How we read your site
Before any of this: our audit reads the file your server actually sends — the raw HTML, with no JavaScript run. That’s what ChatGPT’s and Claude’s crawlers read. Google’s crawler will run the code eventually. Theirs never does. If the audit can’t find your phone number, neither can they.
Raw response, no JavaScript run
HTTP/2 200 content-type: text/html; charset=utf-8 <!doctype html> <html lang="en"> <head><title>Propane Delivery</title></head> <body><div id="root"></div></body> </html>
The twelve, on one screen
Six of these get talked about, several of them incorrectly. Six don’t get talked about at all, and those six are the local ones.
The six you’ll hear about
It’s an app, not a website
Most AI website builders don’t build websites. They build JavaScript applications. The content isn’t in the page. It’s in code that has to run first.
Here’s the part almost nobody says out loud. Google’s crawler will at least try. It indexes the raw HTML first, pulls the links out of it, then puts the page in a queue to be rendered later, usually seconds, sometimes minutes. So Google gets there eventually. Slowly, and at a cost to how much of your site it bothers to crawl.
The AI crawlers don’t try at all. Vercel and MERJ measured more than 500 million GPTBot fetches across Vercel’s network in a single month, and in the same study found that GPTBot does not execute JavaScript — not slowly, not later, not at all. Anthropic’s crawler behaves the same way. They read the file your server sends and that’s the end of it.
Two things to be straight about, because most people selling this get them wrong.
Gemini and Google’s AI Overviews ride Google’s crawler, so they do run the code. You are not invisible to all AI. And the fix isn’t “rebuild it as a static site from 2005.” The requirement is simply that your content is in the HTML your server sends. Server-side rendering does it. So does a pre-built page. So does incremental rebuilding. Only the build-it-in-the-browser approach fails.
What PIMS does
Every page we build serves its content, its headline, and every internal link in the first response from the server. We test it the way a crawler would: request the page with no JavaScript, count the words, count the links. If the raw version isn’t carrying at least 80% of what a person sees, it doesn’t ship. And when we audit a site we didn’t build, that check is the first thing we run on it.
What a person sees
Home · Delivery · Tanks · Service area · Contact
Propane delivery, next day in most of the county
Keep-full and will-call routes, tank sets, and 24-hour emergency service. Get on the schedule and we’ll confirm your first fill.
Serving the county and the towns around it, six days a week.
Schedule a deliveryWhat a crawler receives
<!doctype html> <html lang="en"> <head> <title>Propane Delivery</title> </head> <body> <div id="root"></div> </body> </html>
https://www.example.com/propane-delivery · illustrative capture. example.com is reserved for documentation, so no real company is named.
The plumbing nobody looks at
Two small files decide whether search engines can work with your site at all, and both get shipped wrong constantly.
robots.txt tells crawlers what they’re allowed to look at. We have seen sites launch with the “discourage search engines” box still checked from the staging build. One line, and the whole site is closed for business. Worse: if that one file returns a server error, crawling stops for the entire site, not just that page.
Your sitemap is the list of pages you hand to search engines. Most builders write it once at launch and never touch it again. You add a service page, nobody gets the memo. And the “last modified” date on each entry only helps if it’s honest. Google has said plainly that it only uses that date when it’s consistently accurate, so a sitemap stamping every page with the date of your last deploy is worse than one carrying no dates at all.
What PIMS does
Your sitemap regenerates every time you publish, update, or delete a page, and it carries the real modification date or no date at all, never a fake one. It lists only pages that actually return a live, canonical, indexable response. Your robots.txt gets read in the audit, and it is the first item on our launch gate — one bad response on that one small file stops crawling of the whole site, so it is the first thing we look at, not the last. We also reconcile the count: if your site has 61 published pages and the sitemap has 54, something is stranding live pages and we go find it.
The two files
- robots.txt
- read in the audit
5xx here stops the host - sitemap.xml
- regenerated on publish
lastmod real, or absent
indexable URLs only
Your homepage lives at eight addresses
With www, without www. With a trailing slash, without. Then double all of that for http versus https. Eight addresses, one homepage, and to a search engine that can look like eight competing pages.
The common advice is that a canonical tag fixes this. It doesn’t, quite. A canonical tag is a hint. The fix is a redirect at the server: one address wins, the other seven send you to it in a single hop, and the winner points at itself. The canonical is your second line of defense. The redirect is the tool.
Same story with deleted pages. When you take a page down, your site has to say it’s gone. Two common mistakes: returning “everything’s fine” on a page that no longer exists, or redirecting every dead page to your homepage. That second one looks tidy and gets reclassified as an error anyway.
What PIMS does
We run the full eight-address test on every site, live, and the pass bar is exactly one address answering and seven redirecting in one hop. Deleted pages return the correct gone-or-missing response. If a page has a real replacement, it redirects there. We check for redirect chains across every place redirects can hide, because more than one redirect list is how chains get built without anyone deciding to build one.
http / https · www / no www · slash / no slash · all of it to one canonical address
The machines don’t know what you are
A person lands on your site and gets it in two seconds. Bobtail, phone number, 24-hour emergency service, the county you cover. Done.
A machine is parsing code and guessing. Propane dealer? Cylinder exchange? A blog about propane? Schema is how you stop the guessing. It’s a block of facts in your page’s code that no visitor ever sees: your exact name, address, phone, hours, services, and service area, labeled so a machine can’t misread them.
Now the honest part, because schema gets oversold in both directions. It is not a ranking factor; Google has said so since 2018. And it does not get you cited by AI. That was actually measured this past May by Ahrefs. 1,885 pages that added schema, against about 4,000 matched pages that didn’t, and the difference in AI citations came out at plus 2.4% in AI Mode and plus 2.2% in ChatGPT. Both positive. Both small enough that you would never see them. Anyone selling schema as an AI-citation lever is selling you something that’s been tested and didn’t work.
What it genuinely does is make machines understand what you are. Google and Microsoft both confirm they use it for that. Without it, a machine assembles a picture of your business out of your marketing copy, and sometimes it gets it wrong.
What PIMS does
Every site we build emits the right set and nothing invented. Your organization, one business record per office with its own identity, a service record on each service page, and breadcrumbs that match the actual page structure. We never publish a FAQ block in code that isn’t really on the page. We never publish star ratings you can’t point to real reviews for. We never publish a price range, because pricing is your business, not markup. And every one of those records is checked against the official validators with zero errors before the site goes live.
Measured: schema against AI citations
- pages with schema added
- 1,885
- matched control pages
- ~4,000
- AI Mode citations
- +2.4%
- ChatGPT citations
- +2.2%
Ahrefs, May 2026
The words don’t match what anyone types
Your page title is your label in search results. Your headline is the first thing a reader sees. AI builders write both for a brochure.
“Solutions for your home comfort needs” is a nice sentence. Nobody has ever typed it. What people actually type is propane delivery [town], propane near me, propane tank refill [town], emergency propane delivery, [town] propane company.
And the first paragraph has one job: let the visitor finish what they came to do. What you deliver, where you deliver it, how to get on the schedule. Right at the top, before the mission statement and the company history.
Here’s the reason, and it isn’t the reason you’ll usually hear. You’ll see people claim Google watches how visitors behave on your page and rewards you for it. Google denies that, and we don’t repeat it. The real reason is simpler and better. Somebody just moved into a house, walked out back, read the placard on the tank dome, and is trying to find out who fills it and how fast. If your first paragraph answers that, you get the call. If it opens with your founding year, they hit the back button and call the next dealer.
What PIMS does
One page, one question. Titles written to the search, not to the brochure. One real headline in plain language. The first forty words answering what, where, and how to hire. And these are written by people, not generated at volume. Every word a customer reads goes through our editorial review before it goes anywhere near your site.
What people actually type
propane delivery [town] propane near me propane tank refill [town] emergency propane delivery [town] propane company
Your pages are a pile
AI builders link everything to everything through the nav bar. That makes a pile.
The usual correction is to build a strict tree, one parent, no cross-links. That’s wrong too, and it’s wrong in a way that costs propane dealers specifically. Your Gilbert page and your Chandler page are neighbors. Your propane delivery page and your tank installation page are related. A rigid tree pretends none of that is true. What you want is a neighborhood.
What actually kills a set of town pages is two things nobody mentions. Orphans: town pages that exist in the sitemap and are linked from nowhere on the actual site. And junk anchor text: forty links that all say “learn more,” which tells a machine nothing about where any of them go.
What PIMS does
A hub for services, a hub for your service area, and town pages that link up to their hub, across to the service they’re about, and out to two to four genuinely adjacent towns, chosen by real distance and not by which names look similar. Every page within three clicks of your homepage. Zero orphans, and we test for that by crawling the finished site and looking for any page with nothing pointing at it. No “click here” anywhere.
left, a pile · right, a neighborhood
And that’s where the usual advice stops.
Those six are the ones you’ll hear about. They’re real and we fix all of them. They’re also the half that has nothing to do with propane.
A propane company doesn’t get found by being technically correct. It gets found by being the obvious answer to who delivers to my address. That’s the other six, and it’s where almost all of the money in this vertical actually is.
The six nobody gets to
Your name, address and phone don’t match themselves
Not “roughly match.” Match. The version in your footer, the version on your contact page, the version in your site’s code, and the version on your Google Business Profile have to be the same string of characters. “Ste 4” in one place and “Suite 4” in another is a mismatch a machine can’t resolve.
What PIMS does
We normalize and compare all four, character by character, and any difference fails. Then we check the same thing against the listings that aren’t Google — Apple Business Connect, Bing Places, Yelp, your state propane association directory — and report the drift.
Your service area is a sentence. It should be a circle.
Most propane websites say something like “proudly serving the tri-county area.” A machine can’t do anything with that, and neither can a customer 44 miles out who fills out your form and then finds out you don’t come that far.
You run roughly 30 to 35 miles from each bulk plant, because that’s where the gallons per mile stop covering the truck, the driver, and winter dispatch. That’s a real circle with a real center, and it belongs in your site’s code as a circle.
What PIMS does
A real geographic circle around each office’s actual coordinates, set at that office’s real radius, 30 to 35 miles for most plants, published in the code where machines read it. One per office, never one giant circle drawn around everything. Every generated town page is tested against that circle, and out-of-radius towns don’t publish. If you run outlying towns on a weekly set day, we say so on the page instead of pretending they’re next-day. And your Google service areas are built from the same circle rather than from wishful thinking, which is what stops the out-of-area lead you were never going to serve anyway.
one circle per bulk plant · 30 to 35 miles · boundaries generic, not a named operator
Nobody real is on the About page
“Our team of experts” is a machine-generated sentence and it reads like one. Meanwhile you have an actual credential set most dealers never publish: your years in business, your state license, your association membership, the certifications your techs actually hold, the counties you actually cover.
What PIMS does
A real named About page with the real credentials, and our domain reviewer enforces it both directions. We never claim a certification you don’t hold, and we never let a real one you do hold get quietly dropped because a writer didn’t know it was worth mentioning.
The contact form that silently fails
It works. Probably. Has anybody checked? This is the least glamorous item on this page and the most expensive one. A form that posts successfully, shows the thank-you message, and delivers the lead to nobody looks identical from the outside to a form that works. We have watched it happen. It is silent, and it can run for months.
What PIMS does
Before your site goes live we send a real submission through the form and confirm a human received it. Not a success code, an actual message in an actual inbox. After that we watch the delivery record on every site we run, so a form that quietly stops delivering shows up as a flag instead of a quiet month. Same for the phone number: it has to be real, reachable, and the same everywhere.
Speed, and photos of your actual trucks
Slow pages lose the customer standing in a cold kitchen. We hold the standard thresholds: main content loaded in 2.5 seconds, response to a tap under 200 milliseconds, and no layout jumping around while it loads, measured on real visits rather than on a lab test.
And the pictures. We use real photos. If we’re building on your existing photography, we swap the logo and the phone number and that’s it. We do not feed a real photo of your truck into an image generator and publish whatever comes back. Every image gets honest alt text, which is also the only version of that text a machine ever reads.
The AI crawler list changed again
The list of AI crawlers changed again this year. Anthropic retired one agent name and added two. If your site is still allowing the old one, you’re allowing a crawler that no longer exists and blocking two that do.
And there’s a widespread belief that blocking Google-Extended keeps you out of AI Overviews. It doesn’t. AI Overviews pull from Google’s regular index, through the regular Google crawler. The controls that actually matter are the snippet settings, and a badly set snippet limit, usually inherited from a template, will quietly hold you out of the answer.
What PIMS does
We keep the crawler roster current across every place it’s written down, and we re-check it against Anthropic’s and OpenAI’s published rosters. And we check the snippet settings in both places they can hide: in the page, and in the server response.
Where the snippet setting hides
<meta name="robots" content="max-snippet:-1"> X-Robots-Tag: max-snippet:-1
What we refuse to sell you
You’re going to be pitched all four of these. We’ve tested them. Here’s where we stand.
What you’ll be sold: “llms.txt is cheap insurance.”
It’s a plain-text summary of your site aimed at AI systems. In the study that measured what happened next — 137,000 sites, Ahrefs, 2026 — 97% of published files received zero AI crawler requests. No engine has committed to using it. We’ll drop the file on your site because it costs nothing, but we will never put it on an invoice, and we will never tell you it did anything.
llms.txt study, Ahrefs, 2026. 137,000 sites. 97% zero requests.
What you’ll be sold: “Schema markup gets you cited by AI.”
Measured by Ahrefs at roughly zero this past May, against a matched control group. Schema earns its place because it makes machines understand you. That’s the whole claim, and it’s enough.
Schema-versus-citation intervention study, Ahrefs, May 2026.
What you’ll be sold: “Block Google-Extended to control AI Overviews.”
AI Overviews draw from the regular search index. Blocking that one agent does not remove your page from AI answers.
AI Overviews serve from the regular Google index.
What you’ll be sold: “Google watches how people behave on your site.”
Google denies it. We’re not going to build your first paragraph on a mechanism the company in question says doesn’t exist. We build it on the customer who needs propane today.
Google denies the mechanism on the record.
If a vendor is charging you for any of the four, that’s worth knowing before your next invoice.
The standard
None of the twelve is a judgment call. Every one is a written check with a stated test and a pass-or-fail answer, and they run in order, because there’s no point improving your page titles on a site that’s still telling crawlers to go away.
Ten phases. The twelve problems above all live inside them, in the order they have to be fixed.
- Can crawlers reach it at all
- One address per page, redirects clean
- Content present in the HTML your server sends
- Sitemap accurate, honest dates, Search Console verified
- Name, address, phone, Google profile and service area all agreeing
- The full schema set, correct and nothing invented
- Internal links: hubs, neighbors, zero orphans
- Titles, headlines and openers that match real searches
- Speed, images and mobile
- Social cards, favicon, crawler roster, monitoring
A site doesn’t get called live until it passes. That standard is a real document, and we rebuild it against current dated research rather than against habit. The most recent rebuild was this month.
We wrote the checklist down. It’s a real, dated document — ask any vendor to show you theirs.
Get your free AI-visibility audit
We’ll run the first pass of it on your site.
The score
The checklist is what we do. The score is what you watch.
Start with the free audit. Company name, website, ZIP. Add your email and we’ll send you the report. You get back a plain-language report: where AI assistants currently point people who ask about propane in your county, the specific gaps on your site, a count of the things you can fix, the exact code for the ones that are copy-and-paste fixable, and a 30-day plan. The audit hands you the work, not a grade. Somebody in your county is typing “who can swap out my tank” this week — the report tells you whether the machines send him to you. No sales call required to get it.
Then it keeps score. As a PIMS customer you get an AI Visibility Score from 0 to 100 on a dashboard, updated every week, with a short written note on what moved and why. One number, the changes underneath it, and what we did about them. We don’t hand you a wall of charts.
And you get the county picture. We track where AI assistants are pointing people in your service area, and where they aren’t pointing anyone in particular yet. Those gaps are openings. A question in your county that nothing local currently answers well is the cheapest position you will ever take. We frame it that way on purpose. Nobody is beating you here. This is credit you already earned, from machines that don’t know you exist yet.
One honesty note on the numbers, since you’ll see it in the report: AI visibility figures are a representative sample via web search, not a direct measurement. Engine labels share one web-search reading. Nobody has a direct measurement inside those products. We’d rather give you one number we can stand behind than five that look impressive.
The AI-visibility module is part of PIMS. The audit is free. There is no separate product to buy.
Your score, once you’re on the system
0 to 100, one number, updated weekly on your dashboard, with a written note on what moved. The free audit comes back with your county standing and a list of fixes, not a score.
AI visibility figures are a representative sample via web search, not a direct measurement. Engine labels share one web-search reading.
Why us
Dennis South runs Southwest Propane out of three locations in New Mexico. He’s a founding partner, and his site is the one we point at. The baseline we started from, the weekly scores since, and the Google activity log are all published on his case study page.
The plainest proof we have is the one you’re standing in. PIMS runs Propane Insider’s own marketing: the pages, the newsletter, the search work. If you found us through a search, that was the system doing the thing this page is describing.
Bill Stomp started running one propane location and grew to 87 branch offices across 12 states over a 25-year career, combat veteran. He’s since helped 200+ fuel companies.
More on what the system covers: what PIMS actually does.
What operators ask us
“My site looks fine to me.”
It does. It looks fine to every person who visits it. The crawler receives a different file than the one you’re looking at. If you want to check it yourself in thirty seconds: open your homepage, view the page source, and search that source for your own phone number. If it isn’t in there, it isn’t in the file the AI crawlers get either.
“I already paid somebody to build this site.”
Then the audit costs you nothing and tells you whether you got what you paid for. If it comes back clean, that’s a real answer and we’ll say so. It isn’t a pretext for a rebuild. A fair number of these are fixable without touching the design at all.
“Isn’t SEO mostly stuff nobody can prove?”
A lot of it is. That’s exactly why the refusals section above exists. Four things the industry sells, four measurements, four things we won’t charge you for. We’d rather be the vendor that tells you what doesn’t work.
“I don’t have time for a marketing project.”
You have time for a ZIP code. That’s the whole ask on the audit. Everything after that is our job.
Find out what a machine actually sees.
Company name, website, ZIP code. About sixty seconds. You’ll get where you stand in your county, the specific gaps, the code to fix them, and a 30-day plan. Free, no obligation. Add your email and we’ll send you the report.
Worth doing now rather than in November: every week between here and fill season is a week of “propane near me” searches in your county — and you don’t yet know whether you’re in the answer.
Get your free AI-visibility audit
Rather talk first? Book 20 minutes and we’ll walk through your site together.