How Do I Get ChatGPT to Recommend My Roofing Company? The Five-Slot Problem

01Citation and recommendation are different events

Almost every roofing company asking this question has already done the thing that produces citations and still cannot see itself in an answer. The reason is that a citation and a recommendation are produced by different parts of the same process, and the first does not imply the second.

A citation is a source attached to an answer. A recommendation is your company name appearing inside the sentence a homeowner reads. In an analysis of AI answers, 61.7% of citations were "ghost" citations: the domain was used as a source and the brand was never named. In the same body of work, ChatGPT cited brands in 87% of appearances but named them in only 20.7%. A separate tracker put it at roughly 3.2 times more mentions than links.

For a roofer the commercial consequence is exact. If your cost page supplied the number in "replacing a roof in Markham runs $9,800 to $16,400", you did the work and received no phone call — because the sentence the homeowner read contained no name. Ghost citations are the most expensive failure in this discipline, and they are invisible on every dashboard that counts links.

This is the same distinction that separates a citation play from a recommendation play in our roofer's GEO playbook, and it matters because the two require different budgets. Citation work is page work. Recommendation work is identity and presence work, and it mostly happens off your website.

02How few slots actually exist

The number that should reframe your expectations: in a study that captured AI answers for roofing across 60 US cities, the answers named an average of 4.6 companies per city, never fewer than three and never more than six. Across all 60 cities, 277 distinct companies held those slots.

Put differently: a metro area with thousands of licensed roofers produces a recommendation shortlist of roughly five names. Everything else is invisible to that buyer at that moment. The same study found 59 of 60 cities where no single roofing company owned the answer.

A second dataset supports the shape from the opposite direction. Across 90 answers in five US metros, one study found AI recommendation openness running 59 to 67 out of 100 for roofing, which it described as a smaller set of businesses repeating more often than in the other trades it measured. In Tampa, 55 businesses were named across 18 answers, with one leader named in 10 of 18. And on two of six questions, the assistants named no business at all.

That last figure is worth sitting with, because it is sometimes read as "AI will not recommend anyone". It is not. It says that AI recommendation is selective compared with the Google local pack — 1.2% against 35.9% — and that selectivity is precisely what makes a slot valuable. The roofer in the answer gets a disproportionate share of the buyers who ask the question.

03The four states, and what each is worth

Before the tactics, a taxonomy. A roofing company in an AI answer can be in one of four states, and knowing which one you are in determines the fix.

Two useful properties of this taxonomy. It is diagnosable from outside — a monthly prompt log tells you which state you are in without vendor tooling. And it is sequential: you cannot jump from state one to state four, because each state's fix is the entry condition for the next.

The implication for budget is uncomfortable. Most roofing marketing spend targets state one, because that is what SEO sells. The leads live in states three and four, which are bought with listings, mentions and proof rather than with pages — which is why the roofing SEO programme has to be paired with identity work to reach them.

04Why an assistant decides to name anyone

An assistant names a business when naming it is the cheapest way to answer the question. That framing is more useful than any list of ranking factors, because it tells you what to build: evidence that resolves to a name.

Four conditions have to hold at once. Remove any one and you drop to a lower state.

• The entity has to resolve. The system must be confident that the company in a directory listing, the company on a review site, the company on a manufacturer locator and the company on a website are one business. Conflicting phone numbers, an old address or a name variant breaks that resolution, and an unresolved entity is not nameable.
• The category has to fit. Roofing prompts get roofing answers. A general contractor with a roofing page competes badly against a business whose every signal says roofing — which is why category tags on listings matter more than the description field.
• The geography has to match. "Near me" is resolved from location and from text. If no readable source connects you to the city in the prompt, you are not a candidate for it.
• There has to be corroboration. One source saying you serve Markham is a claim. Five independent sources saying it is a fact. The corroboration layer is where the rest of this article lives.

Our AI SEO case study walks through the same resolution problem from the search side: the work that moved visibility was entity and corroboration work, not page production. The pattern is consistent across clients.

05Branded mentions: the strongest lever measured

If you want one number to organise this entire discipline around, it is 0.664. That is the correlation between branded web mentions and AI visibility reported by Ahrefs in its 2026 analysis — the strongest single correlate the company measured, ahead of backlinks.

Read the direction of that relationship carefully. It is not about people linking to you. It is about people writing your name in sentences. A mention inside a local news article, a manufacturer's certified-contractor page, a supplier's case study, a trade association listing or a genuinely useful forum thread is the raw material an assistant uses to justify naming you.

The asymmetry is the point. The bottom row is what gets sold to roofers because it is easy to sell. The top four rows are what gets you named, and three of them are free if you are already certified and already buying materials.

There is a related finding that explains why mention work outperforms link work in this channel: in one dataset, 86% of citations came from sources a brand owns or manages — split between first-party websites at 44% and business listings at 42% — while ChatGPT still drew 48.73% of its citations from third-party sources. You control the first-party and listing layers, which is the argument for consolidating them before chasing coverage you cannot control.

The practical weekly routine is small and unglamorous: ask every supplier for a case study after a major job, claim the association listings you are paying for, and answer one forum question a month as a named local professional rather than as a brand account. That is a realistic programme for a roofer with no marketing department, and it addresses the strongest lever in the data. Our roofing local SEO strategy collection expands this into a fuller list.

06The recommendability dossier

Here is the artefact this article exists to give you. Six fields, each one verifiable by a stranger in five minutes, and each one a condition for being nameable. If any field fails, that is your next task.

Score yourself honestly on each: pass, partial, fail. The distribution usually tells the story — most roofers pass field one partially and fail fields three and five entirely, which is exactly the combination that produces ghost citations rather than mentions.

None of this is exotic, and that is the point. It is the same local-search hygiene that has been recommended for a decade, and it has become the entry ticket to a channel where the shortlist is five names long. The search engine optimization work for roofing professionals that used to be optional housekeeping is now the difference between being nameable and not.

07The lists that are the recommendation

For roofing specifically, the recommendation layer is largely pre-built by other people. In a 600-answer ChatGPT roofing baseline, at least 82% of answers cited a directory, review site or "best of" list, and only 17.8% cited a roofer's own website. The lists are the answer.

Cheers Market Baseline, roofing questions, 31 August–20 September 2026, 600 ChatGPT answers. Perplexity shares from the same dataset. A separate Florida study of 745 queries ranked BBB first at 272 citations, ahead of Angi at 170 and Yelp at 160.

Two operating conclusions follow.

Audit only what is cited in your market. The national shares are a starting point, not a task list. Run the prompt panel in section twelve, record which domains appear in your city, and audit those. A roofer in Denver spending a day on YellowPages.ca has wasted a day, and a roofer in Mississauga who has never claimed HomeStars has a hole in the exact place ChatGPT looks.

Treat a listing as a page, not a form. The fields that matter are the ones a machine reads: name, category, phone, hours, service area, licence, rating. The description field matters far less than roofers assume, and the photo gallery matters least of all. This is the opposite of how listing sales are usually pitched, which is why it is worth saying plainly.

There is one more layer specific to this trade: manufacturer contractor programmes were named by Gemini in 56.3% of roofing answers. If you are certified with a major manufacturer, your locator entry is a recommendation asset. If it carries a dead number, it is a liability that no amount of website work offsets. And the paid counterpart of this layer — ChatGPT Ads for roofers — reaches buyers who never scroll to a list, which is worth having as a separate line rather than as a substitute.

08Cross-platform presence: the unclaimed wedge

The most exploitable finding in the roofing data is also the simplest. In a study of 745 queries across 10 Florida metros and five platforms, no roofing business appeared on all five. Of the roofers seen named, 97.2% appeared on only one platform.

That is a market failure, and failures are opportunities. It means the businesses being recommended are mostly accidental winners of one engine's particular fetish — a strong BBB record, a well-populated Expertise profile, a homepage that suits AI Mode — rather than companies deliberately present everywhere.

The wedge is not clever tactics. It is consistent presence across all four reward structures at once, which is achievable in a quarter and which almost no competitor will do because it involves four boring jobs rather than one interesting one.

This is the layer where an AI SEO programme differs from conventional SEO: the deliverable is not a ranking, it is the same business resolving identically across four systems that select sources differently.

09Trust markers that roofing uniquely requires

Roofing carries a fraud and trust concern that plumbing and heating do not, and the research reflects it. In the Florida study, BBB was the most-cited source for roofers at 272 queries, ahead of Angi at 170 and Yelp at 160, and the authors attributed that directly to trust and scam concerns in the trade.

That has a practical consequence for what you publish. A homeowner asking an assistant about a roofer after a hailstorm is not only asking "who is good" — they are asking "who will not disappear with my deposit". The markers that answer the second question are the ones that make you nameable.

• Licence number, written out. On the site, in the footer, and in the listing's licence field where one exists. A plate number is the cheapest authenticity signal available.
• Insurance and bonding, stated as a sentence. "Fully insured, certificate available on request" is a claim; naming the coverage type and stating the certificate is available before work starts is evidence.
• Accreditation status, accurate. If you are BBB accredited, say so, and make sure the BBB record agrees. If you are not, do not imply it.
• A written warranty you actually honour. Publish the term, what it covers, and what voids it. Vague lifetime language is a red flag to a homeowner and unquotable to an assistant.
• A local address, not a mailbox. Roofers operating from a virtual office score badly on corroboration, and it shows up as an inability to resolve the entity to a place.
• Named people. An owner's name and face is a trust signal in a trade where the failure mode is abandonment mid-job.

None of these is a ranking factor in any documented sense. They are the conditions under which an assistant has something concrete to say about you when a homeowner is worried — and being the company with something concrete to say is the whole game in states three and four.

The commercial mirror of this is worth noting: the same trust markers reduce wasted sales time, because they pre-qualify. That is why the high-intent roofing lead work and the recommendation work share the same inputs.

10Ratings are a floor, not a ladder

Roofers are sold review volume as the path to AI visibility. The published data does not support the ladder version of that claim, and it supports the floor version emphatically.

The median named roofer has 130 reviews — a fifth of the heating and cooling median — and nearly a quarter of roofing answers name a company with fewer than twenty. Volume is not the gate. What is nearly universal is the floor: 99.4% of rated appearances showed a company at 4.0 stars or above.

One overlooked asset: review replies. They are descriptive prose about your own work, published publicly, in your own words. "We replaced fourteen squares after the August hailstorm and traced the leak to a step-flashing failure" is a fact an assistant can use. "Thanks for the kind words!" is not. Reply to every review, including the bad ones, and describe the work.

11Which prompts produce recommendations

Not every prompt can produce a recommendation, and testing the wrong ones is how roofers conclude that AI search does not work for them.

Two rules follow. Weight your panel toward selection and emergency prompts, because those are the ones where a name is the answer, and track them separately from cost and material prompts where a citation is the realistic ceiling.

And do not read a material-prompt absence as failure. If you are not named for "is a metal roof worth it", you are in the same position as every other roofer and the same position as most manufacturers' dealers. The measurable contest is the selection class, and that is where the five slots are decided.

The measurement framework our team applies to the appointment stage — the point at which a recommendation becomes revenue — is covered in the booked-appointment SEO strategies; the prompt classes above are the front end of the same funnel.

12Measuring share of recommendation

Four numbers describe recommendation, and most reporting measures the wrong one. In order of usefulness to a roofer:

Track the fourth one first even though it is about other people. Knowing which competitors are named, and how often, tells you whether your metro is one of the 59 open ones or the one where a single leader holds 10 of 18 answers. Those two situations call for different strategies: in a concentrated market you are fighting for one slot; in an open market you are competing against instability itself.

# Roofing recommendation panel — 20 prompts, logged, monthly
# Logged out. Private window. Same day each month. Same order.
#
# prompt | engine | me named? | my position | competitors named (in order) | sources cited
#
# Selection class (weight these):    8 prompts, city-modified
# Emergency class:                   4 prompts
# Claim class:                       4 prompts
# Cost class:                        4 prompts
#
# Read monthly: recommendation share, average position, competitor occupancy
# Read quarterly: cross-platform coverage, and whether the same names persist

A warning about sample size, because it is where most measurement goes wrong. Same-day repeats of an identical question have been reported to overlap by only about 41% in domain terms. One screenshot is noise. Twelve monthly runs of a fixed twenty-prompt set is evidence — and that also means you cannot judge a change in under a quarter, which is longer than most roofing clients are asked to wait.

Join the panel to revenue with two things: call tracking, so a call from a directory card is attributable, and an intake question — "how did you first hear about us?" with an AI assistant option. BrightLocal's 2026 survey puts AI use for local business recommendations at 45%, up from 6% the year before. Some of those homeowners are already calling you and being credited to organic search.

13Why storm season compresses everything

Roofing has a demand pattern other trades do not: a weather event creates a ten-day window in which thousands of homeowners ask the same question, and the answer they receive determines who they call. In that window, recommendation share converts to revenue at a rate no other month will match.

It is also the worst possible moment to begin. The mechanisms are all slow: entity resolution, listing correction, mention accrual and cross-platform coverage each take weeks, and the reference answers in a storm window are being assembled from sources that have existed for years. A page published the morning after a hailstorm is competing against a directory profile that has been citable since 2019.

One prompting detail that surprises roofers: emergency prompts often produce an unnamed answer. In the five-metro study, on two of six questions the assistants named no business at all, and emergency phrasing is prone to producing advice rather than a list. That is not a failure of your optimisation; it is a property of the question. Track emergency prompts for contact-detail presence rather than for rank.

14What stops you being named

• A tracking number in a listing. The most common single cause of a broken entity. One mismatch between your site and your BBB record is enough to make you unresolvable.
• Two trading names. "Example Roofing" on the site and "Example Roofing & Exteriors Ltd." on a directory splits your identity in half, and half an identity does not earn a slot.
• Content aimed at being cited. Perfect if you want a footnote. If you want the name, the content has to say the name — and third parties have to say it too.
• Buying listicles. Paid "best of" placements are increasingly discounted, and they consume the budget that would have produced a supplier case study that actually corroborates.
• Chasing review volume. With a median of 130 reviews among named roofers and 23.6% of answers naming companies under twenty, a thousand-review programme is solving a problem the data says you do not have.
• Ignoring Canadian directories. HomeStars appeared in 40% of ChatGPT's Canadian answers and none of its US ones. A Canadian roofer auditing American directories is working the wrong list entirely.
• Reading a single check as a result. With roughly 41% same-day overlap, one screenshot proves nothing in either direction — including the screenshot that shows you in the answer.
• Promising a client a position. Nobody can guarantee a slot in a generated shortlist. What can be verified is entity consistency, listing accuracy, mention coverage and cross-platform presence.

15The 90-day plan to become nameable

Realistic expectations at day ninety: identity and listing fixes change which state you are in before they change your volume, and in a five-slot market the first movement is usually from absent to cited, not from cited to named. Naming follows mentions, and mentions accrue on a slower clock — which is the honest reason this is a two-quarter programme rather than a campaign.

16Sources and confidence

Every figure above traces to one of the following. Most of the trade-specific data comes from agencies and research vendors with a commercial interest in the conclusion, so it is labelled rather than presented as independent research.

How this page is built for machine reading. A summary block above the first section, a direct answer in the first hundred words, one idea per paragraph, an explicit question-and-answer module, defined terms at first use, and structured data for the article, the FAQ, the glossary and the procedure. Page semantics follow W3C HTML guidance, markup uses the public schema.org vocabulary, and every figure is inline SVG rather than a late-loaded image, keeping the page free of external requests in line with the Core Web Vitals guidance.

Four caveats. The five-slot figure comes from a single engine (Google AI Mode) on a single day, so treat 4.6 as an order of magnitude rather than a constant. The roofing shares rest on two questions in a four-week window. The 97.2% single-platform figure is from one state, one month. And no published study tests whether a mention programme produces more booked roofing jobs — the chain from mention to name to call is a well-evidenced inference, not a measured result, and it should be validated in your own market before it becomes a budget line.

17Questions we get asked

18Glossary

Recommendation

A business named inside the answer text, as opposed to a source cited beneath it. Only a recommendation reliably produces a call.

Citation

A source attached to an answer. Can occur without the brand being named anywhere in the sentence.

Ghost citation

A citation where the domain is used and the brand is not named. Reported at 61.7% of citations in one analysis.

Neutral reference

A business described in an answer without being endorsed as a recommendation — present, not shortlisted.

Recommendation share

The share of tracked prompts in which your company is named. The core metric of this discipline.

Position

Place within the named list. Home-services data puts the average at 2.9.

Openness

How dispersed recommendations are in a market. Low openness means a few businesses repeat; high means the answer is still unsettled.

Entity resolution

A system's ability to determine that a website, a directory profile and a locator entry refer to one business. Broken by name or phone mismatches.

Branded mention

Your company name written in prose by a third party. Correlates at 0.664 with AI visibility — the strongest single measure reported.

Corroboration

Independent sources agreeing on the same fact about you. One source is a claim; five is a fact.

Recommendability dossier

The six-field self-audit in section six: identity, category, geography, proof, claim and contactability.

Cross-platform coverage

How many distinct engines name you at least once. 97.2% of named roofers in one study had coverage of one.

Prompt panel

A fixed set of prompts asked in the same order monthly, logging names, positions, competitors and cited sources.