I Asked 4 AIs to Recommend a Roofer. 12 of 13 Got No Link.

In August 2026 I ran the same test on thirteen roofing companies. Five questions a homeowner would actually type, put through four AI assistants, scored on one thing: did the answer hand the homeowner a link to the company's own website?

Twelve of the thirteen scored 0 out of 5. The thirteenth scored 1.

13

roofing companies tested across eleven states and one national franchise

260

answers read: five prompts, four assistants, thirteen companies

12 of 13

companies never had their own website linked in any answer

1

link to a tested company's own site across all 260 answers

45

mentions of a shingle manufacturer, the name the assistants gave most instead of a local roofer

This is not a story about bad roofers. Some of these companies have strong Google rankings and deep review histories. It is a story about what these engines can see, and about what they quietly build their answers out of. Below is the whole thing: method, raw results, and the five patterns that explain them.

How I tested 13 roofing companies in 4 AI assistants

Five prompts per company. Not keyword strings, but the questions people actually ask an assistant when the ceiling is dripping. The wording was written for each company's market and services, so no two companies got identical prompts, but every set was built from the same five shapes:

1. 01Best [service] in [city or state], 2026. The plain who-should-I-call question, in the company's own market.
2. 02Is [competitor] worth it? or Reviews of [competitor]. A homeowner vetting a rival by name.
3. 03[Competitor A] vs [Competitor B]. A homeowner comparing two other local companies, to see whether anyone else gets pulled in.
4. 04Where to find [specific service] in [area]. Gutters, storm damage, metal, commercial, multi-family; whichever the company actually sells.
5. 05How to choose a roofing contractor. The generic question every homeowner asks first.

For example, O'LYN Roofing's five were: best roof replacement services in Greater Boston, 2026; where to find reliable gutter repairs in Norwood; are GF Sprague's roofing services worth it in 2026; Colonial Exteriors vs. New England Gutter & Exteriors, which is better; and how to choose the right roofing company. Each prompt went to four engines: ChatGPT, Perplexity, Gemini and Claude. That is 20 answers per company and 260 answers in total, collected on August 21, 2026.

Scoring was deliberately narrow. A prompt counted as a win only if at least one engine put a link to the company's own website in its answer, as a recommendation the homeowner could click. Not a passing name-drop. Not a directory page that happened to list them. And not an engine quietly reading the company's page in the background and then leaving it out of the answer. That happened to three companies in this test, and it is logged separately below. Their website, in the answer, as a link. That distinction turns out to be the most important thing in the entire study, and I will come back to it.

The results: 12 of 13 roofing sites never got a link

Twelve companies scored 0 of 5. One, O'LYN Roofing in Greater Boston, scored 1 of 5. Here are all thirteen, with the names the engines handed out instead.

The summary table is available as a spreadsheet: download the study data as CSV, 13 rows, one per company, with the score and the names each assistant gave instead. The answer-by-answer records for five of the companies are published too; the data appendix at the end says what is in them and what was not kept. Use it, quote it, check it.

Shamrock Roofing produced what I would call a clean zero: across all 20 answers, no site citations and not a single mention of the brand name. The engines filled a Kansas City roofing question entirely with DaBella, Erie Home, Aspen Contracting and GAF: three national consolidators and a shingle manufacturer.

Five patterns that decide who ChatGPT and Perplexity recommend

1. Manufacturers and consolidators own the answer

GAF, a shingle manufacturer that does not install anything, showed up as a top recommended name for nine of the thirteen companies. Tecta America, CentiMark, DaBella, Erie Home, Leaf Home, Power Home Remodeling and Aspen Contracting took most of the rest.

That is not a quality judgment. Those brands have national footprints and enormous volumes of citable content, so they are the safest thing for an engine to reach for when it has nothing local and specific to hold onto. If your market's answer is a wall of national brands, the engine did not find a local source it trusted. It found a vacuum.

2. Directories and review platforms are the wiring

Look at what the answers were assembled from and the same handful of domains appear over and over: BBB, Angi, Expertise.com, HomeGuide, Thumbtack, HomeAdvisor, Birdeye, Trustpilot. In the Elite Roofing sample, BBB alone was cited 12 times and Angi 9 times.

Reviews are part of this, not decoration. In Phoenix, Phoenix Roofing & Repair's own site was cited five times, and Perplexity justified its pick with the company's 4.9-star average across 1,056 reviews. Review count was the reason an engine gave for a pick more often than anything else, and it travels far outside Google.

3. Your competitor wrote the list you are being judged on

This one changed what I would publish. In Wichita, a blog post titled around best roofing companies in Wichita was cited in four separate answers. It was published by Premier Roofing, a competitor. A roofer wrote the ranking that the engines then used to rank roofers.

It kept happening. In Tyler, the engines cited a competitor's storm-damage service page (Arrowhead Roofing) five times, plus a directory called serviceagent.ai. In Tampa, blog posts were the single largest source type at 54 citations. For Elite Roofing, blog posts were the largest source type too, at 36 citations.

The best roofers in [city] listicle that AI treats as neutral authority is very often written by a roofer who is on it.

4. Ranking on Google does not carry over

Trust Roofing in Tampa is the cleanest proof in the dataset. They have 2,647 ranking keywords, 72 of them in the top 10, and 181 references in Google AI Overviews. By any traditional SEO scorecard they are winning.

They still scored 0 of 5. No llms.txt, no FAQ schema, an off-spec meta description. Nothing there makes a page easy for an assistant to lift a clean, attributable answer from. Meanwhile Elite Roofing in Mississippi has FAQ schema, which most roofers skip, but only 3 ranking keywords and zero AI Overview references, so there is nothing for the schema to help. You need both: something worth citing, and a page shaped so it can be cited.

The bluntest example: Claude pulled eliteroofer.com in as a source on a Mississippi roof repair question, read it, and still gave the homeowner no link. The most it did was list Elite Roofing & Gutters thirteenth among the names in that answer. It was not alone. Rhoden Roofing's site was read on the Wichita prompt, and O'LYN's roof replacement page was read by Claude on the Boston prompt, and in both cases the engine named the company without linking it. Getting read is not the same as getting recommended.

5. Getting named and getting linked are not the same thing

Rhoden Roofing scored 0 of 5 on links, but the brand name did surface on the residential roofing prompt. Gemini ranked them first, Perplexity first, Claude second. ChatGPT left them out entirely. Istueta Roofing in Miami was slipped into two of Gemini's lists, never first and never with a link, and the other three engines did not mention them at all.

A name with no link is a reputation win and a traffic loss. The homeowner reads the name, then has to go look you up on their own. A linked answer sends them to your page with your phone number on it. Track them separately or you will congratulate yourself for the wrong thing.

Why O'LYN was the only one to score

O'LYN Roofing in Greater Boston officially scored 1 of 5, and the official number undersells them. All four engines named O'LYN on both local prompts. Three of the four listed them first for the roof replacement question, and Claude read their roof replacement page while doing it. On the Norwood gutters prompt, Perplexity and Gemini read their gutters page too. ChatGPT went further and put a link to that gutters page in its answer. That was the only click any engine sent to any of the thirteen companies, and the one that put a point on the board.

The detail worth stealing is which page got linked. Not the homepage. A specific service page, about one job, in one region. That is what an engine can hand to a person as an answer. GF Sprague (5 mentions) and Roof Hub (4) are right behind them in that market, so the lead is thin. But it exists, and nobody else in the study had one.

What I would do to get a roofing company recommended by AI

The full playbook is in how to get recommended by ChatGPT. The short version, in the order I would do it:

1. 01Run the test on yourself first. Five prompts, four engines, twenty answers. Write down who wins in your market and which sources the answers were built from. That list is your to-do list.
2. 02Build one real page per service and per city you serve, with specifics: crew names, job photos, permit and code details, price ranges, response times. Generic pages give an engine nothing to quote, and these are the same pages that carry roofing SEO in ordinary search, so the work pays twice.
3. 03Add a genuine FAQ block, written as questions with short direct answers. That is the shape engines lift.
4. 04Fix the plumbing: llms.txt, FAQ schema, a clean meta description, consistent name, address and phone across every listing.
5. 05Get into other people's content. Directory profiles, supplier and manufacturer contractor listings, local press, association pages, and yes, your own city roundups, because your competitors are already writing them.

One caveat I will state plainly: these engines do not give the same answer twice. Run the same prompt tomorrow and the ordering will shift. What does not shift is the shape: national brands, directories, review platforms and competitor blogs filling the space where a local contractor's own pages should be. That held across all thirteen companies, from a Boston residential roofer to a national franchise.

If you want to know where your company lands, request the free test and I will run the same five-prompt test on your business and send you the raw answers, free, with no pitch attached. You can also work through the free 31-point website audit yourself in an afternoon; most of what caused these zeros is on that list. If you want the test re-run every month with the fixes done for you, that is what the AI search visibility work is. And if it turns out the fix is a real site rather than a patch, I build those at a fixed price quoted in writing before anything starts: a five-page site or a full build with a page per service and town, live in one to two weeks.

Data appendix

Two files sit behind this article. The summary has 13 rows, one per company. The answer-level file has 100 rows, one per answer, for the five companies whose full records survive: Elite Roofing & Gutters, Trust Roofing, Istueta Roofing, Perry Roofing and Whitaker Roofing.

For the other eight (Coryell, Stonewater, Five Guys, Rhoden, Shamrock, O'LYN, Summit Point and Storm Guard) I kept the scores and the names given instead, but not the answer-by-answer records. They were working files, deleted before I decided to publish the evidence, and I have not rebuilt them. The details above about those eight, including O'LYN's linked gutters page and Rhoden's rankings by engine, were written from those records at the time and cannot be checked against a published file.

How it was collected

• When: August 21, 2026. The five surviving records were collected between 12:21 and 14:46 UTC.
• How: a third-party visibility-testing tool, called through its API. It writes the prompts, sends each one to the four assistants with no account history and no location, and records what comes back.
• Models, as recorded: ChatGPT gpt-4o-mini, Perplexity sonar, Gemini gemini-3.1-flash-lite, Claude claude-haiku-4-5. These are the smaller API models, not necessarily what the consumer apps answer with.
• Prompts: written by the tool from each company's website, five per company. Some name a competitor ("is X worth it", "X vs Y"), and some are state-wide or national. The tool tags each one tofu, mofu or bofu (early research, comparing, ready to hire); the tag is in the file.
• One run per company. Nothing was repeated, so the file is one day's answers, not an average.

What the columns mean

• own_site_linked: the tool's flag that the answer put the company's own website in front of the reader. It is the study's score: a prompt counts if any of its four answers has it.
• company_named: the company appears in the tool's list of brands named in that answer, link or not.
• cited_domains and cited_urls: the sources the tool recorded for the answer, the pages the assistant drew on. A source is not a click; nothing here measures whether a homeowner followed anything.
• answer_excerpt: the first 400 or so characters of the answer. The tool stores only that much, so the full answers were never kept.

Checks against the published numbers

• Recomputed from the answer file, all five companies score 0 of 5, as published.
• The tool's own headline counter recorded no mentions for any of the five, but its per-answer brand lists name Istueta twice (both Gemini) and Elite Roofing & Gutters once (Claude, thirteenth in the list). The company_named column uses the per-answer lists.
• The names-given-instead counts in the summary are the tool's totals. Counting answers instead can come out one lower, because a name can appear twice in one answer: Watertight Roofing is 4 in the summary and appears in 3 answers.
• Four sources are pages from marketing firms and are shown as [marketing firm, withheld]; I don't name other agencies. Nothing else is edited.

Both files are free to use under CC BY 4.0, with credit to Riptide Growth.