How Roofers Get Cited in ChatGPT: A Technical GEO Playbook

01The uncomfortable baseline

Start with the number that should govern your whole strategy. It comes from the Cheers Market Baseline, a three-week collection running 31 August to 20 September 2026 across 60 US and 40 Canadian metros — 34,742 usable answers in total, of which 600 were roofing answers from ChatGPT.

Across those 600 roofing answers, ChatGPT cited a roofing company's own website in 17.8% of them. For the same period, the same method found ChatGPT citing a contractor's own site in 44.0% of HVAC answers and 53.7% of plumbing answers. Roofing is not plumbing with a different keyword set. It is a structurally harder citation market, and the reason is in the next paragraph.

At least 82% of those roofing answers cited a directory, review site or "best of" ranking list. The distribution: Expertise.com 40.0%, Angi or HomeAdvisor 34.5%, HomeStars 25.7%, BBB 20.8%. In other words, when a homeowner asks ChatGPT for a roofer, the assistant is mostly quoting aggregators — and your company appears only if an aggregator already lists you, or if your own site survives a much narrower filter.

• 17.8%of ChatGPT roofing answers cited a roofer's own websiteCheers, 600 answers, Aug–Sep 2026
• 82%+of roofing answers cited a directory, review site or "best of" listCheers, same sample
• 1.2%of brand locations ChatGPT recommends at all, across 350,000+ locationsSOCi Local Visibility Index 2026
• 45%of consumers now use AI for local recommendations, up from 6% a year earlierBrightLocal Local Consumer Review Survey 2026

Two more context numbers keep this honest. First, demand is real: BrightLocal's 2026 Local Consumer Review Survey records local-recommendation use of AI rising from 6% to 45% in a year, with ChatGPT leading at 31% and Google AI Mode at 23%. Second, supply is thin: Roofing Contractor's 2026 survey found only 11% of roofing contractors using AI search services to generate leads, even as 40% report using AI somewhere in the business. A large and fast-growing discovery channel, and almost nobody in your trade competing on it yet.

One caution before the mechanics: a citation is not a lead. BrightLocal's research also found that only 18% of consumers who used AI for local search felt ready to contact the business it recommended — the rest kept researching elsewhere. AI visibility wins the shortlist, not the phone call. The paid route into the same conversations is covered separately in ChatGPT Ads for roofers, and the evidence on whether that paid route converts is in our Google Ads vs ChatGPT Ads comparison.

02Five gates between you and a citation

Most roofing "AI SEO" advice collapses five distinct technical gates into one vague instruction: publish helpful content. That is why it does not work. A page has to clear all five, in order, and each fails for a different reason.

• 1

  Access — can the crawler fetch you at all?

  Governed by robots.txt per user-agent, firewall rules, and whether your pages render without JavaScript. Fail here and nothing downstream matters, because the page never enters an index.

• 2

  Indexation — are you in the corpus that retrieval searches?

  ChatGPT search draws on OpenAI's own web index with fallback to established search infrastructure. A page that is not indexed cannot be retrieved, regardless of how good the answer block is.

• 3

  Retrieval — do you make the candidate set?

  Roughly 16.6 URLs are pulled per prompt in Ahrefs' 1.4-million-prompt study. Relevance to the reconstructed question decides this, not keyword match — a query fan-out of sub-questions, not one string.

• 4

  Selection — is your chunk worth quoting?

  Retrieved is not cited. Zyppy's 2025 analysis found roughly 15% of retrieved pages get cited. What survives is a self-contained, evidence-bearing passage that answers the question without needing the surrounding page.

• 5

  Rendering — do you appear as a name or a card?

  The final gate. Answers often render as a structured list of businesses; Suff Digital found card lists in 96.8% of ChatGPT answers to "best roofer in [city]" questions. Being named requires structured, consistent business facts, not prose alone.

Work them in order and you will find that most roofing sites fail at gate one or gate three, and spend all their effort on gate four. Our overview of AI search optimization is organised around the same sequence for that reason.

03The four OpenAI crawlers, and the one line that matters

This is the most consequential and least glamorous section in the article. Per OpenAI's published bot documentation, it runs separate user agents for separate jobs, and they are controlled independently in robots.txt. Getting this wrong produces a silent, total failure that no rank tracker will show you.

The failure mode is a wildcard. A roofing site whose host added User-agent: * with a broad Disallow — or a firewall challenge applied to bots — blocks OAI-SearchBot along with everything else, and vanishes from ChatGPT answers while continuing to rank normally in Google. You will see no error, no drop in Search Console, and no explanation.

The configuration that keeps citations while making a deliberate training decision looks like this:

# Allow ChatGPT search and user fetches — these drive citations
User-agent: OAI-SearchBot
Allow: /

User-agent: ChatGPT-User
Allow: /

# Optional: opt out of model training only
User-agent: GPTBot
Disallow: /

# If you run ChatGPT Ads, keep landing pages readable
User-agent: OAI-AdsBot
Allow: /

Sitemap: https://yourdomain.ca/sitemap.xml

Three operational details worth putting in the job ticket rather than the strategy deck. OpenAI documents that a robots.txt change takes roughly 24 hours to reflect in search behaviour. Each agent publishes its own IP ranges — openai.com/gptbot.json, openai.com/searchbot.json, openai.com/chatgpt-user.json — so you can verify in server logs that the traffic you think is arriving actually is. And when fetching robots.txt itself, the agent appends a robots.txt marker to the user-agent string, which lets you distinguish a policy fetch from a content fetch when your log format drops the path.

04Retrieval is not citation

The distinction that separates working GEO from content marketing with an AI sticker on it. Two studies bracket it from different directions.

Ahrefs analysed 1.4 million ChatGPT prompts and found the system pulled roughly the same number of URLs it did not cite as URLs it did — about 16.6 of each per prompt. One striking illustration from the same study: Reddit is fetched through a dedicated channel, yet was cited for under 2% of the Reddit URLs retrieved.

From the other side, Zyppy's 2025 research concluded ChatGPT cites only about 15% of the pages it retrieves. And Zhang, He & Yao's 2026 analysis of 21,143 citations found ChatGPT averages 6.88 citations per prompt — fewer than Google's 12.06 and Perplexity's 16.35 — while attributing markedly more influence to each one it does cite (mean fetched-page influence 0.2713 against 0.0584 for Google and 0.0646 for Perplexity).

Read together, those findings have a precise implication for a roofer: ChatGPT is a low-breadth, high-conviction engine. It cites a handful of sources and leans on them hard. So the goal is not to be mentioned everywhere at low quality. It is to be one of the two or three sources an answer cannot be written without.

There is a matching implication for how you write. If a single cited source carries disproportionate influence, then a page that says one thing precisely beats a page that says six things vaguely — which is the opposite of the "comprehensive ultimate guide" instinct most roofing content follows. The deeper mechanics of how such answers are assembled are in our analysis of ChatGPT's matching behaviour, which applies equally to organic citation.

05Chunk-level citability, and what the research actually says

This is where received wisdom and published evidence disagree, and roofers deserve to know which is which.

What the evidence supports. Zyppy's 2025 analysis found 44.2% of ChatGPT citations come from the first 30% of a page. Whatever your most quotable claim is, it needs to be near the top — not after a 400-word brand introduction. Aggarwal and colleagues' 2024 GEO paper, presented at KDD, found adding quotations, statistics and citations lifted visibility by up to about 40% in benchmark conditions. Zhang, He & Yao's 2026 observational analysis found evidence genres — definitions, statistics, comparisons — produced +57% to +77% higher mean influence than format changes alone. And Axios's October 2026 coverage of citation research noted practical guides accounting for 68% of citations when answering how-to questions, buyer's guides cited in 45% of "best-of" answers, and expert analysis leading industry-trend questions at 41%.

What the evidence does not support. The same Zhang study found that Q&A formatting alone did not improve absorption — a −5.74% relative difference across 23,745 citation features. In other words, converting your headings into questions is not a lever. What gets absorbed is evidence: a number, a comparison, a definition, a source. A roofer who writes "What does a roof replacement cost?" and then answers with three adjectives has done the formatting and skipped the substance.

The writable version of this lesson: write the paragraph you want quoted, put it in the first 30% of the page, and put a number in it. Everything else on the page is supporting material for that paragraph. Our longer treatment of this, with fifty specific asset ideas, is the 50 AI SEO strategies for roofers.

06The directory layer is the actual battleground

If 82% or more of ChatGPT's roofing answers cite a directory, then directory presence is not a legacy local-SEO chore. It is the primary citation surface for your trade. The question is which directories, because the answer is not the same in Canada as in the United States.

Note the Yelp figure and let it sink in: 0.6% of home-services answers. Canadian contractors in particular have spent a decade on Yelp presence that the citation data does not reward at all. Meanwhile a directory most roofers have never claimed — Expertise.com — leads the roofing set. There is a second-order finding worth flagging, because it changes what "listing accuracy" means: Yelp appeared as a citation in only 0.6% of answers, yet in one five-day capture 34% of the rated business cards on ChatGPT's map linked to a Yelp page. A source can be unimportant as a citation and important as a data provider. Both matter; they are different jobs.

The practical sequence for a roofing company, in priority order:

1. Claim and correct every directory that actually appears in your market. Name, address, phone, service list, hours, licence and insurance status — identical everywhere. Inconsistency is what removes a business from consideration before a citation is even possible.
2. Add the ones that lead your vertical, even if they cost money and even if they were never on your radar. Expertise.com's 40.0% share in roofing answers is the clearest single fact in this article.
3. Then fix your own site — because a directory listing usually links somewhere, and if that somewhere is a thin page, the referral converts badly.

This sequencing is the reverse of what most agencies sell, which is the point. The roofing SEO company page describes how we order it, and the trade-specific constraints are covered on SEO for roofing professionals.

07The Canadian directory gap

Canada-specific

This section exists because almost every published GEO study is American, and applying it to a Canadian roofer produces the wrong work order. The same Cheers dataset split its answers by country, and the result is one of the most actionable findings available to a Canadian contractor.

Across the full 32-question set, ChatGPT cited HomeStars in 40% of Canadian answers and in none of the US answers. Conversely, Angi and Expertise.com were effectively absent from Canadian answers — the two directories that dominate the American roofing set. If your agency has sent you an audit built on American directory priorities, it is measuring the wrong list, and a Canadian roofer could spend a year on Expertise and Angi placement while the actual gap — HomeStars — goes unworked.

Two further Canadian specifics change the plan. Roofing answers in Canada frequently involve insurance claims and provincial code language, and answers that name a manufacturer contractor programme (GAF, Owens Corning, CertainTeed and similar) appeared in 56.3% of Gemini's roofing answers and 31.3% of Google AI Mode's — so manufacturer programme membership is a citation asset in this trade, not merely a lead-gen badge. And if you work both sides of the border, you need two directory stacks and two content sets, because the retrieval corpus differs by country.

Market-specific context for the metro you serve: Florida roofing and Colorado roofing differ from Canadian conditions, and within Canada the equivalent pages are Toronto contractor marketing and Vancouver contractor marketing. For the general contractor-wide version of this playbook, see SEO for contractors and SEO strategies for booked appointments in Gemini and ChatGPT.

08Which page earns the citation

Roofers are routinely told to blog more. The citation data says blog posts are close to irrelevant to this particular outcome.

When ChatGPT cited a contractor's own website in the roofing set, the destination broke down as follows: homepage 37.4%, service page 24.1%, service-plus-city page 17.4%, location page 10.7% — and blog posts 1.2%. Of the four engines measured, when any of them cited a roofing company's site, 67.7% of those citations went to the homepage.

That is a striking inversion of SEO habit. In classic local SEO the homepage is often the least-editable page and the blog is where the effort goes. In AI citation, the homepage is the single highest-yield surface, because it is the page that most reliably carries the entity facts — what you do, where, since when, licensed how, rated what — in a form that can be lifted.

The honest nuance: blog posts earn only 1.2% of contractor-site citations, yet practical guides take 68% of citations for how-to questions. Both are true because the how-to citations mostly come from other domains — publishers, forums, reference sites. So your blog's job is narrower than you were told: it is a supporting asset that occasionally wins a how-to answer, not the engine of local visibility. The engine is your homepage and your service pages. This is the same conclusion we reached empirically in our AI SEO case study.

09Entity, earned media and manufacturer programmes

Here the published figures conflict, and the honest move is to show the conflict rather than pick the flattering number.

Estimates of how much of ChatGPT's citation diet is Wikipedia range widely: 47.9% of top factual citations in 5W's 2026 index, about 12.1% of citations in Analyze AI's 83,670-citation study spanning November 2025 to January 2026, and 7.0% of share in Ahrefs' Brand Radar ranking published 2 September 2026. Those are not reconcilable as measurements of one thing — different query sets, different denominators, different definitions of a citation. What is consistent across all of them, and across Semrush's 2026 AI Visibility Index of 126 million US prompts, is the direction: earned and reference sources carry the weight, and brand-owned pages are a minority of what gets cited. One frequently quoted figure from 5W puts 85.5% of AI citations on earned media rather than brand websites.

For a roofing company, "Wikipedia" is mostly not available — notability is not on the table for a regional contractor, and self-editing is a bad idea regardless. So the earned-media target for a roofer is different, and three channels are realistically open:

1. Manufacturer contractor programmes. Gemini named a shingle manufacturer's contractor programme in 56.3% of its roofing answers. Being listed, accurately, on GAF/Owens Corning/CertainTeed locators is a citation and entity signal in a trade where those brands carry authority the roofer does not.
2. Trade and local press. A named expert quoted on storm-damage documentation or a provincial code change is earned media, and expert analysis leads citation of industry-trend questions at 41%. This is where a real owner's name outperforms an anonymous brand.
3. Community presence, handled honestly. Reddit is heavily cited in aggregate (16.8% of ChatGPT citations per Ahrefs' Brand Radar, 2 September 2026), and is also the most volatile source on the platform — Semrush tracking shows ChatGPT citing Reddit in close to 60% of responses in early August 2025 and about 10% by mid-September. Participate; never manipulate. Peec AI found roughly one in ten AI citations between December 2025 and February 2026 came from self-promotional listicles, and that engines are actively tightening against them.

Entity clarity is the quiet prerequisite for all of it. The same legal name, address, phone, category and service list must appear identically on your site, in your schema, on every directory, and in every earned mention. Inconsistency is not a minor blemish — in a system that assembles an answer from multiple retrieved pages, a business described three different ways is three weak signals instead of one strong one. Our roofing marketing engagements start with that reconciliation for exactly this reason, and the same logic is why we treat brand and entity work as part of AI SEO rather than as a separate brand exercise.

10Reviews as a citation and rendering input

Roofers already chase reviews for the local pack. The AI layer gives reviews a second, separate job — they decide whether you render inside the card list at all.

SOCi's 2026 Local Visibility Index, covering more than 350,000 locations across 2,751 brands, found that locations ChatGPT recommended averaged 4.3 stars. The card-level data from 2026 collections is more encouraging than the doom-laden framing suggests: the median review count among roofers named was 91 for ChatGPT cards, 130 for Google AI Mode, and 152 for Gemini cards — counts a regional company can reach. Ratings were a firmer floor: only 0.6% of roofers named by AI Mode showed under 4.0 stars, and 84.1% showed 4.8 or higher.

Layer BrightLocal's consumer thresholds on top and the target becomes concrete: 68% of consumers will only use businesses with 4+ stars (up from 55% a year earlier), 47% will not use a business with under 20 reviews, 74% only care about reviews from the last three months, and 89% expect the owner to respond. That last one is the cheapest win available: an owner who replies to reviews is satisfying a stated consumer expectation and manufacturing fresh review signal at the same time.

11What schema does, and what it does not

Schema markup attracts the most confident advice and the least consistent evidence in this field, so here is the awkward data.

Ahrefs tracked 1,885 pages that added JSON-LD between August 2025 and March 2026 against roughly 4,000 matched control pages. The measured effect on citations: −4.6% in AI Overviews, +2.4% in AI Mode, +2.2% on ChatGPT. That is a rounding error in two cases and a small negative in the third. Against this, Authoritas research reported that pages with FAQ schema and inline citations were weighted roughly 40% higher, receiving about three times more citations than equivalent unstructured prose — a much larger claim, from a different method, on a different engine mix.

The defensible reading: schema is not a citation lever; it is an ambiguity reducer. Its real function for a roofer is to state, unambiguously and machine-readably, who the business is, where it operates, what it sells, what it charges and who vouches for it — so that a system assembling an answer from several sources resolves your business as one entity rather than three. It also powers the card layer, which is where 96.8% of "best roofer in [city]" answers put their names.

So implement it, for the right reason, in this order: LocalBusiness (or the RoofingContractor subtype) with areaServed, address, telephone, openingHours, priceRange and sameAs; Organization with logo and sameAs links to every profile; Service on service pages; FAQPage only where the questions and answers are genuinely on the page; Person for the owner with credentials; and Review/AggregateRating only where policy permits and the reviews are real. Do not expect a visible citation lift from any of it, and do not let a vendor sell you schema as an AI SEO fix.

For the implementation depth on entity markup, the podcast episode on ranking in Google AI Overviews, Gemini and ChatGPT covers the technical checklist, and the outlook for advanced roofing SEO in the AI era covers where this is heading.

12Measuring without fooling yourself

The most important measurement fact in this article: when researchers asked ChatGPT the same questions twice on the same day, only 41% of cited domains overlapped. A single screenshot of an answer is noise. Any agency presenting one as a benchmark is either measuring badly or selling.

Three distinctions make a measurement system usable. First, mentions, citations and clicks are three different scores: an answer can name your brand without linking your site, link your site without naming the brand, or do both — and none of them is a phone call. Second, compare within a fixed panel, never across ad-hoc prompts. Third, attribute the source layer: record whether the citation went to your domain, a directory, or a publisher, because that tells you which of the playbook's sections to work next.

The panel we run for roofing clients uses four query classes, run weekly, with every cited URL logged:

Then connect it to leads, or do not bother. Track AI-referred sessions in analytics as their own channel, watch assisted conversions rather than last-click, and remember the BrightLocal finding that only 18% of AI-assisted local searchers felt ready to contact the business — most continue researching. An AI citation is a mid-funnel event with a long tail, and reporting it as if it were a lead source will get it cancelled in month two.

13The 90-day build

Sequenced by dependency, not by ease. Nothing in week eight works if week one is skipped.

Weeks 1–3, access and reconciliation. Confirm OAI-SearchBot and ChatGPT-User reach your pages in server logs; fix robots.txt and any WAF challenge; make the homepage and service pages render server-side. In parallel, build one canonical record of your business — legal name, address, phone, services, licence, hours — and diff it against every existing listing. Most roofers discover four or five variations.

Weeks 4–6, listings. Claim and correct the directories that actually appear in your market, Canada-first. Add the vertical leaders. Add manufacturer contractor programme locators. Re-verify after a fortnight, because third-party directories drift.

Weeks 7–9, citation surfaces. Rebuild the homepage top-third as an entity and definition block. Rewrite each service page so the problem, constraint, price band and timeline appear in the first 30% of the page, in prose that survives being quoted alone. Add service-plus-city pages for the two or three areas that actually produce revenue.

Weeks 10–12, evidence and measurement. Pitch two data-led stories to trade or local press. Institute review response and a velocity target. Stand up the twenty-prompt weekly panel and record the baseline before you change anything else, so that month four's improvement is attributable rather than assumed.

If you want the paid channel running alongside this while citations compound, that is the roofing PPC conversation, and the interactions between the two — including the eligibility constraint that ads only reach Free and Go users — are in the ChatGPT Ads for roofers playbook.

14What not to do

Six ways roofing companies waste money and credibility on this channel, each with the reason attached.

1. Do not publish llms.txt. Google Search ignores it by its own statement, and in May 2026 97% of published llms.txt files received no requests at all. It is a format with no demonstrated distribution.
2. Do not block OAI-SearchBot to "opt out of AI". That is GPTBot's job. Blocking the search agent removes you from the answers while achieving nothing for training policy.
3. Do not cloak crawler traffic. Serving different content to bots than to users is both a rejection reason on the paid side and, at the landing-page level, a fast route out of the retrieval index. It is also the specific pattern that gets sites deindexed from AI surfaces.
4. Do not write self-referential listicles. Peec AI measured roughly one in ten AI citations between December 2025 and February 2026 coming from self-promotional "best tools" listicles, and found engines tightening against them. A roofer's "Top 10 roofing companies in [city]" featuring itself first is the same pattern at a smaller scale.
5. Do not keyword-stuff to compensate for low citation rates. Aggarwal's GEO research measured keyword stuffing at roughly −8% visibility. The counter-programming is quoted evidence, not keyword density.
6. Do not buy a "guaranteed ChatGPT ranking". There is no such mechanism, no report to verify it against, and the answer set changes enough that same-day repeats agree only 41% of the time. Anyone selling it is selling you a screenshot.

In AI search, the honest measurement is a small number you tracked yourself for twelve weeks — not a screenshot someone else chose for you. Ahmet Dogan, CEO, LeadGulls Digital Marketing Agency

15Sources and confidence

Every figure above traces to one of the following. Where studies disagree, both figures appear in the body of the article and the disagreement is stated rather than resolved in the flattering direction.

Where a figure is vendor-published, it is labelled vendor. Where two credible studies disagree, the article states both. If you know of Canadian roofing citation data we have missed, we would like to see it — this register is reviewed quarterly.

16Questions we get asked

17Glossary