01What high intent means when the query is a sentence
In paid search, intent is a property of a keyword. "Roof leak repair near me" is high intent; "how long does a shingle roof last" is not; you bid accordingly, and the auction tells you the price of the difference. In AI search there is no keyword and no auction. There is a prompt, often twenty words long and containing three facts about the asker's house, and an answer that names two to eleven businesses.
That changes the unit of optimisation twice. First, you cannot buy placement per query, so you have to earn inclusion in the answer's candidate set. Second, inclusion is worth very different amounts depending on where in the answer you land — and the research on that is new enough that most roofing marketing plans have not absorbed it.
The practical definition we use with contractors: a high-intent AI query is one where the answer ends the research phase. It names businesses, implies a next action, and leaves the reader with nothing further to learn. A roofing answer that lists three companies and their licence numbers is high intent. A roofing answer explaining ice damming is not, no matter how many words it shares with a money query.
Your competitor is not the roofer in position two. It is the homeowner's assumption that the first name in the answer is the safest one to call, at 11pm, with water coming through a ceiling.
That assumption is doing the selling. Your job is to be legible enough — licensed, reviewed, correctly listed in the directories the assistant reads, with an answerable phone — to be one of the names it picks.
02The economics of a position, not a visit
First Page Sage measured how often ChatGPT users actually click through to the brands named in an answer, across 31,470 responses recorded in September 2026 by an opt-in panel of 2,150 US adults, combined with referral data from 112 websites across ten industries. For home services, the position gradient is steep.
ChatGPT click-through rate by position, home services, September 2026
Two conclusions follow, and they are not the ones the AI-visibility industry sells.
First, moving from third to first roughly multiplies expected clicks by 2.4. Not citations, not impressions — clicks and calls from people who were going to hire someone in the next hour. If you are running a roofing operation where an installed replacement is your profit centre, that is the highest-value position in marketing, and it is decided by which businesses the assistant trusts enough to name.
Second, 8.4% of users skip the click entirely and go search the brand name. For a roofer this is not a rounding error; it is likely the single largest AI-driven lead pathway, and it will show up in your analytics as organic brand search, or as Direct, or as nothing at all. Section 04 is about getting that traffic counted.
03The honest evidence on AI conversion rates
You will be sold a multiplier. Every AI SEO pitch deck in 2026 contains one, usually Seer's 15.9% for ChatGPT against 1.76% for organic. Before you build a business case on it, look at the whole evidence base — because the most rigorous studies in it disagree.
| Study | Sample | Finding | Limit |
|---|---|---|---|
| Amsive, 2025 | 54 websites, six months of GA4 data, paired analysis | 4.87% LLM vs 4.60% organic — p = 0.794, not statistically significant. B2B subset did show a gap: 2.17% vs 1.16%. LLM traffic was 0.24% of sessions | High site-level variance; small LLM session counts |
| Seer Interactive | One B2B client, mixed verticals | ChatGPT 15.9%, Perplexity 10.5%, Claude 5.0%, Gemini 3.0% vs Google organic 1.76%; ChatGPT sessions viewed 2.3 pages | Single client, B2B; not transferable to roofing |
| Adobe Analytics, Q1 2026 | Over one trillion visits, 130+ North American retailers | AI referral converted 42% better than non-AI in March 2026, reversing from 38% worse a year earlier; +37% revenue per visit | Retail; vendor-stated, not independently audited |
| Marketing Science (Kaiser & Schulze) | 973 e-commerce sites, ~50,000 ChatGPT transactions vs 164M other transactions | ChatGPT converted 13% below organic and 32% below paid search | E-commerce transactions only |
| Shero Commerce, 2026 | 73 stores, 96.4M sessions, 976,668 orders | AI-referred visits converted at 1.27% vs 1.09% organic — a 25% lift at median, far below headline claims | E-commerce; median, so tail effects are invisible |
| Lebesgue, April 2026 | 35,000 e-commerce brands on Shopify | AI referrals converted 3.6% vs 1.23% for traditional search, close to brand-query performance at 3.65% | DTC brands |
| First Page Sage, 2026 | 160+ client companies, May 2025–Jul 2026 | ChatGPT conversion by industry: HVAC 4.0%, construction 3.6%, pest control 3.6%, solar 3.3%, legal 5.4%, hotels 7.2%, e-commerce 3.2%, B2B SaaS 2.5% | Client base of one agency; definitions of "conversion" vary by client |
The most useful reading is not that one side is right. It is that the premium concentrates where the buyer has a research phase — legal, hotels, higher education, B2B — and disappears where the transaction is a commodity checkout. Roofing sits awkwardly in the middle: a homeowner researching ice damming is doing research; a homeowner whose ceiling is leaking is not. Those two arrive through the same assistant and convert nothing alike.
Do not put a 15.9% conversion assumption into a roofing forecast. The largest well-controlled test on this question found no significant difference, and the sample sizes in AI referral segments are small enough that a single good month looks like a channel effect.
Use the range as a hypothesis about relative quality, then measure your own segment for two full quarters before you hold it to a target.
04Why the click is not the lead
Roofing leads leak out of analytics in three places, and every one of them is bigger in AI than in classic search.
Referrer stripping. AI applications do not always pass a referrer header, and ad blockers and browser privacy settings remove it from a large share of sessions regardless. Industry analyses put the share of AI visits that land in the "Direct" bucket at 20–40%, with some practitioners reporting 60–70%. Those are vendor estimates without a shared methodology, so treat the range as directional — but the mechanism is real and independently verified by anyone who has watched a client's Direct traffic grow without a campaign behind it.
Brand-search follow-up. The 8.4% figure in section 02 is the one to internalise: nearly one in twelve home-services users read the answer, remembered the first name, and went to search for it rather than clicking. In your analytics that is an organic brand query ten minutes later. If you do not separate brand from non-brand in reporting, you will attribute it to "SEO is improving" and never fund the thing that produced it.
The phone. Contractors overwhelmingly field high-intent roofing demand as calls, and a call from a mobile browser has no session to attribute. In paid search this is solved with dynamic number insertion; in AI search the same tool is the only way you will ever see the lead. If you are tracking forms only, you are blind to the majority of the leads this channel produces.
Where an AI-driven roofing lead goes after the answer
The discipline that follows is blunt: do not judge this channel by sessions. Judge it by inbound calls, brand-search lift in your service area, and jobs you can attribute through a source field on the intake form. Everything else is instrumentation bias.
05Five prompt classes that produce roofing jobs
Not all roofing prompts are worth engineering for. Sort them by what the answer does and what the asker needs next, and five classes emerge. Each has a different surface, a different page, and a different conversion action — so a single "AI SEO" workstream that treats them identically will waste most of its budget.
| Class | What the answer does | Commercial value | What it tests on your site |
|---|---|---|---|
| 1 · Emergency "leaking roof right now", "emergency tarp tonight" |
Names two or three businesses with hours and a phone number; often skips links entirely | Highest. Decision is today; price sensitivity is lowest; job value is a repair that frequently converts into a replacement | Whether your hours, emergency window and phone are machine-readable — and whether anyone answers at 11pm |
| 2 · Cost "what should a roof replacement cost in [city]" |
Gives ranges, then names two or three contractors; cites a pricing page if one exists | High. Buyer is budgeting, often pre-insurance-decision | Whether you publish a range with a date on it, or say "call for a quote" |
| 3 · Selection "best roofing company in [city]", "who should I hire" |
Returns a card list of businesses, usually built from directories and review platforms | High, and the class where directory presence dominates | Your listings, review count and recency — rarely your website copy |
| 4 · Technical "is this ice damming or a flashing failure" |
Explains, with citations, and usually names no business at all | Low direct, high downstream. The diagnostician who answered is the one they call in March | Whether you have field-experience content the assistant can quote |
| 5 · Insurance "will insurance cover hail damage to my roof" |
Explains process and documentation; sometimes names contractors who handle claims | Medium-high. Longer cycle, larger ticket, more documents | Whether your claims process is described as a process, with what you provide adjusters |
Two things about this distribution matter for budgeting. Classes 1, 2 and 5 are where the jobs are, and all three are transactional or near-transactional — the intent band where AI Overviews barely appear and zero-click rates collapse. Class 3 is the biggest single roofing surface and is won almost entirely off-site. Class 4 is the class most agencies optimise because it is the easiest to demonstrate: you can show a citation in a screenshot. It is also the class least likely to produce a phone call this month.
The zero-click asymmetry is worth stating precisely, because it is the most useful strategic fact in this article. Transactional searches are zero-click about 31% of the time and informational searches about 74%; AI Overviews appear on roughly 16.5% of transactional queries against 39.4% of informational ones. High-intent demand is the part of search that AI features have interfered with least. For a contractor, that is the whole opportunity: the money queries still end in a click, a call, or a named business.
06The surface each class runs through
For classes 1, 3 and 5, the answer is usually assembled from sources that are not your website. In the 2026 roofing citation data we reviewed for the GEO playbook, at least 82% of ChatGPT roofing answers cited a directory, review platform or "best of" list; a roofer's own site appeared in 17.8% of roofing answers, against 44.0% for HVAC and 53.7% for plumbing.
That does not mean your site is irrelevant. It means the site's job is different from what SEO taught you. Of contractor-site citations in that dataset, 37.4% went to a homepage, 24.1% to a service page, 17.4% to a service-plus-city page, 10.7% to a location page, and 1.2% to blog posts. The blog is a credibility asset quoted in technical answers; the homepage and service pages are the citation surfaces for commercial ones. A roofing company that spends its AI budget on blog volume is optimising the 1.2%.
Directory layer first, because it decides selection answers. Homepage and service pages second, because they carry the facts an assistant quotes. Reviews third, because they are the filter both of those layers are read through.
Content volume last, and only where it demonstrates field experience. A blog post is not a lead-generation asset in this channel; a diagnostic page that gets quoted into an emergency answer is.
07The page that has to exist
For each prompt class there is one page that has to exist, in a specific shape. This is the build list we hand to roofing clients, ordered by expected lead contribution rather than by effort.
| Page | Must contain, as extractable text | Schema | Converts because |
|---|---|---|---|
| Emergency service page | Response window stated in hours; what you do in the first visit; tarping terms; the after-hours number; the cities you will drive to at 2am | Service + openingHoursSpecification; RoofingContractor for the business | It answers "can someone come tonight" without a form |
| Cost page by job type | Ranges with variables — pitch, layers, decking, eave length — and a visible last-reviewed date | Service with priceRange; FAQPage for the cost questions | A range qualifies the caller before they ring, so the calls that arrive are real |
| Homepage, top third | Business type, licence number, service area as a city list, years under the current entity, review rating and count, the two or three services you want | RoofingContractor with areaServed, telephone, address, sameAs | It is the single most-cited page type in commercial answers |
| Insurance claim page | The documentation you provide, the sequence with the adjuster, what the homeowner must do, and what you will not do | HowTo; FAQPage | It reduces the perceived risk of a claim the homeowner has never made |
| Service-area page per city | Genuine local substance: housing stock, roof types common there, permit quirks, the crews who work it | Service scoped with areaServed | It is what gets retrieved for city-modified prompts, and it is the easiest page to fake badly |
| Diagnostic library | Symptom-first entries in the homeowner's vocabulary, with the faults you actually see, and honest triage on what can wait | FAQPage only where Q&A is genuinely on-page | It earns the technical citation that precedes a spring replacement call |
Note what is absent: no "about us" essay, no gallery page, no service index. Those are support material. The six pages above are the ones that produce calls, and four of the six can be built from facts you already have in a CRM and a licence certificate.
08Conversion mechanics for an 11pm visitor
An AI-referred roofing visitor arrives with an unusual combination: high conviction and low patience. They have already been told who to call. They are checking whether you look real. Every additional step is a chance to lose them to the second name in the answer.
- A tap-to-call number in the first screen, on every page. Not in a sticky bar that loads after the hero, not behind a "contact" tab. The visitor is on a phone, at night, possibly standing under a leak.
- The response window, stated as a commitment. "Calls answered until 9pm, emergencies triaged overnight" is more persuasive than "24/7 service" because it is checkable. If you cannot answer at 2am, say what happens instead — a callback at 7am is a real answer.
- Licence, insurance and warranty above the fold. These are the facts the assistant probably cited. Their absence on your page creates the contradiction that sends the visitor back to the answer.
- A three-field form, not eight. Name, phone, and one selector for the problem. Everything else can be asked on the call. Multi-step forms have their place in paid search where the click was cheap; here the click was expensive and rare.
- An honest price range on the page the answer cited. If the assistant told them "typically $9,000–$18,000 depending on pitch", a page that says "call for a free quote" reads as a retreat.
- No chat widget that pretends to be a person. A bot claiming to be "Sarah from the office" burns the trust the citation just built. If you use one, label it.
- No PDF-only documents. Warranty terms, spec sheets and claim checklists locked in PDFs are invisible to extraction and slow for a phone at 11pm.
Then the operational half, which no amount of SEO fixes. In the 2026 trade surveys we track, 74% of homeowners follow a recommendation and 49% of roofing contractors run Google Ads, while only 11% use any AI search service. The bottleneck for the 11% who do is rarely visibility. It is that the phone rings at 11:20pm and goes to voicemail, or the estimate goes out three days late, and the homeowner who was told you were the safest option calls the second name. Speed-to-estimate is a marketing metric in this channel, and it belongs in the same report as cost per lead.
09Instrumentation: making the lead visible
Four instruments, in order of value. If you install only one, install call tracking; if you install only two, install call tracking and the intake source field, because those survive every platform change while analytics identifiers do not.
1 · The AI referrer segment
Build one GA4 segment rather than hunting per-platform reports. In GA4, create a condition group on Session source matching chatgpt.com, chat.openai.com, perplexity.ai, claude.ai, gemini.google.com, copilot.microsoft.com and you.com. Tag every outbound link you control in directories and profiles with a UTM so at least the traffic you sent yourself is clean.
2 · Call tracking with dynamic number insertion
Provision numbers per channel — organic, paid, AI segment, directory profiles — and let the script swap the displayed number by session source. Two rules: the numbers must forward without a greeting delay a caller will hang up on, and the same number must appear wherever the citation points, or the attribution dies at the moment of the call.
3 · The intake source field
One question on the booking form and one in the call script: "How did you find us?" with options including "an AI assistant". Self-reported attribution is unfashionable and it is also the only instrument that captures brand follow-up, directory card calls and word of mouth in a single field. It costs nothing and it disagrees with analytics in exactly the way that teaches you something.
4 · The CRM stage that matters
Map lead → inspected → estimate sent → won, and get the won value into the same table as the source. Without that last hop you will optimise for lead volume, because lead volume is the only thing your dashboard can see. The moment job value joins the table, "AI search produced 6 leads" becomes "AI search produced two replacements and a repair", and the budget conversation changes.
Because 8.4% of users search the brand instead of clicking, a rise in branded search volume in your service cities is a leading indicator for this channel. Track it by city in Google Search Console and in Google Trends, weekly, on the same chart as your prompt-panel position.
It is a proxy, not proof. It will also rise if you run a radio spot or sponsor a minor-league team. Note those events on the chart rather than pretending the correlation is causal.
10The one number to optimise: cost per booked job
Cost per lead is a vendor's metric. It rewards whoever sends the most leads, regardless of whether they close. Roofing economics are unforgiving about this: a lead that never reaches inspection is a pure loss, and a lead that inspects and declines is a real cost against your sales capacity. Compute in the other direction.
Then compare channels on that basis rather than on CPL. The only external benchmark worth putting on the same page is a paid-search cost per lead, and even that needs a currency and geography label. For reference, the widely quoted 2025 US cross-industry dataset from WordStream/LocaliQ puts Home & Home Improvement at a $8.33 average CPC, 8.05% conversion rate and $90.92 cost per lead — US dollars, US accounts, all campaign types, lead-gen definitions that include form fills nobody qualified. Treat it as an order of magnitude, not a target, and note that Canadian auction pricing generally runs lower outside Toronto and Vancouver.
| Measure | Channel A (paid search) | Channel B (AI search programme) |
|---|---|---|
| Monthly spend | $3,000 | $1,800 |
| Leads | 33 at $90 CPL | 12 at $150 CPL |
| Inspections booked | 18 | 9 |
| Jobs installed | 5 | 4 |
| Cost per lead | $90 | $150 — looks worse |
| Cost per booked job | $600 | $450 — actually better |
The pattern in that table is what we see repeatedly: AI-sourced roofing leads arrive in smaller numbers, at a higher nominal CPL, and with a materially higher rate of conversion to inspection, because the assistant has already narrowed the choice. If you stop the analysis at CPL you will defund the better channel; if you stop at cost per booked job with a sample of four, you will overreact to noise. Hence the next section.
11The 90-day measurement build
Ninety days is the minimum honest window. Amsive's site-level variance was 7.53%, which is enough to make a four-week comparison look decisive when it is not. Sequence it so that the instrumentation exists before the content changes, or you will never know which caused what.
| Week | Work | Artefact you end the week with |
|---|---|---|
| 1 | Freeze a baseline: GA4 AI referrer segment, branded search by city, call volume by source, CRM stage counts. Take a screenshot and store it. | A one-page baseline you will not be able to reconstruct later |
| 2 | Install call tracking with DNI and publish distinct numbers to the site, directories and Google Business Profile. Add the intake source field. | Attribution that survives referrer stripping |
| 3 | Build the 20-prompt panel: five prompts per class (emergency, cost, selection, technical, insurance), city-modified, asked in a fixed order every week. | A weekly position log, not a screenshot |
| 4–5 | Fix the machine-readable facts: RoofingContractor JSON-LD, licence and insurance as visible text, service area as a city list, hours and emergency window. | Six pages that can be extracted without JavaScript |
| 6–7 | Directory corrections in the platforms the assistant actually cites in your market, Canadian or American as applicable. Reviews requested with a description of the work. | A listing layer that states the same facts as the site |
| 8–10 | Rebuild the emergency and cost pages; put the phone, response window and price range above the fold; cut the form to three fields. | Two pages that can convert an 11pm visitor |
| 11–13 | Hold everything still and read the panel, the call log and the brand-search curve together. Compare 30-day windows, not weeks. | A decision: continue, change one variable, or stop |
If you cannot resource weeks 1 through 3, stop. Content work without instrumentation produces a case study you cannot defend when the owner asks which jobs came from it, and roofing owners are right to ask.
12Where AI search beats paid, and where it does not
AI search is not a substitute for Google Ads or Local Services Ads in roofing, and any agency framing it as one is selling against their own numbers. The honest comparison:
| Demand type | Paid search | AI search |
|---|---|---|
| Storm event, next 72 hours | Wins. You can be bidding within hours of a weather alert | Loses. Answer sets and retrieval indices do not move on your schedule |
| Steady-state emergency demand | Strong, expensive, competitive | Strong and cheaper per job once position is earned — but the earning takes a quarter |
| Planned replacement, 3–9 month horizon | Works, with long attribution tails | Wins. This is a research-heavy buyer, the exact profile where the conversion premium is documented |
| Insurance-claim work | Works | Wins when the process is documented on a page the assistant can quote |
| Competitor conquesting | Wins. Buy the auction | No equivalent. You cannot bid on a name in an answer |
| Latency to first lead | Days | 60–90 days minimum, often longer for selection class |
The practical conclusion for a roofing P&L is that these are different instruments for different demand. Paid buys you the storm week and the conquest; AI search compounds the planned-replacement and claims demand where the buyer researches. A roofer who cuts paid search to fund AI SEO will lose the storm week and will not replace it. A roofer who spends everything on paid will keep renting the same clicks every month. Our roofing PPC practice and the AI SEO practice are deliberately separate teams for this reason, and the Google Ads versus ChatGPT Ads comparison sets out what the paid side of AI inventory actually costs today.
13Reviews as an intent filter
In roofing, reviews are not a trust badge. They are the filter an assistant applies before it names anyone, and they are the first thing a homeowner checks after the answer. Two datasets set the thresholds.
What consumers require. BrightLocal's 2026 Local Consumer Review Survey reports that 68% of consumers will only use a business rated 4 stars or higher, up from 55% a year earlier; 47% will not use a business with fewer than 20 reviews; 74% only consider reviews from the last three months; and 89% expect the owner to respond. That last expectation is the cheapest lever available: replying satisfies a stated consumer requirement and manufactures fresh review signal at the same time.
What assistants actually name. In the 2026 roofing citation data, the median review count of a business named in an answer was 91 for ChatGPT, 130 for AI Mode and 152 for Gemini — and 84.1% of named businesses sat at 4.8 stars or above, with only 0.6% under 4.0. SOCi's 2026 index put the average recommended business at 4.3 stars. The two readings are compatible: the assistant does not require perfection, but the businesses it names cluster at high ratings and three-digit review counts.
The engineering implication is unglamorous. A roofing company with 40 reviews averaging 4.6 is not going to win a selection prompt against one with 140 averaging 4.9, and no amount of schema markup closes that gap. Get reviews asked for systematically — with a description of the work, because that language is what gets quoted — respond to every one, and stop buying the five-star badge widgets that add neither count nor recency.
14Five things not to buy
- A guarantee of citations. No vendor controls what an assistant retrieves on a given day, and the same-question-twice overlap in AI answers measures at only about 41% on a single day. Anyone guaranteeing you "top position in ChatGPT" is guessing or lying.
- A conversion multiplier in a forecast. If the deck contains 15.9% and no mention of Amsive's null result, the vendor has read the flattering half of the literature.
- Rank-tracking dashboards as the deliverable. A citation tracker is a diagnostic, not an outcome. Ask what it will be correlated against — calls, inspections, installed jobs — and who owns that join.
- Content volume billed by the article. In contractor-site citations, blog posts were 1.2% of the total. Volume is the wrong unit for this channel; extractable facts on the six pages in section 07 are the right one.
- Directory "verification" packages. Paying a service to mark your listings as updated is not the same as correcting the fields the assistant reads — hours, licence, service area, review count. Do the fields yourself and check them quarterly.
"Which of your numbers is a lead, and how do you know?" If the answer is impressions, citations or visibility scores, the reporting stops one layer short of the only thing a roofing business is buying.
15What we would do first, in order
If a roofing company asked us to start tomorrow with a modest budget, this is the order. It is deliberately boring for the first three weeks.
- Answer the phone. Audit a week of call logs against call volume. If more than a tenth of inbound high-intent calls hit voicemail unreturned within an hour, fix operations before buying visibility. There is no marketing solution to an unanswered call.
- Turn on call tracking and the source field. Two days of work, and it makes every later decision possible.
- Correct the facts. Licence, insurance, service area as a city list, hours, emergency window, price ranges with dates. Visible text first, structured data second.
- Fix the directory layer in your own country. For Canadian roofers this means HomeStars before anything built for the US market; American directory advice transfers badly north of the border, as the Vancouver contractor work keeps demonstrating.
- Rebuild the emergency and cost pages. Phone and response window first screen, three-field form, a real range with a date.
- Ask for reviews with a description of the work, and reply to all of them, weekly, until you are past 100 and current within 90 days.
- Then measure for ninety days and only then decide about content volume, new locations or paid reallocation.
Notice that five of the seven items improve paid search and Google Business Profile performance at the same time. That is the useful property of this work: unlike most AI-specific tactics, the entity and conversion fixes pay for themselves across every channel a roofer runs. The AI SEO case study shows the before-and-after on the entity layer; the contractor SEO service covers the broader technical base.
16Sources and confidence
Every figure above traces to one of the following. Where credible studies disagree — which they do, loudly, on AI conversion rates — both appear in the body and the disagreement is left standing.
| Source | What it supports | Confidence |
|---|---|---|
| First Page Sage, September 2026 | 31,470 ChatGPT responses, panel of 2,150 US adults, referral data from 112 sites in 10 industries; home services CTR 12.1% / 7.2% / 5.0% by position; 8.4% branded search follow-up; legal 13.2% | Published study, agency |
| First Page Sage, ChatGPT Conversion Rates 2026 | 160+ client companies, May 2025–Jul 2026; HVAC 4.0%, construction 3.6%, solar 3.3%, legal 5.4%, hotels 7.2%, e-commerce 3.2% | Agency client base |
| Amsive, 2025 | 54 websites, six months GA4, paired analysis; 4.87% vs 4.60%, p = 0.794; LLM sessions 0.24% of total; site-level variance 7.53%; B2B 2.17% vs 1.16% | Most rigorous controlled test |
| Seer Interactive | ChatGPT 15.9% vs organic 1.76%; Perplexity 10.5%, Claude 5.0%, Gemini 3.0%; 2.3 pages per ChatGPT session; cited brands earn ~2× clicks across 53 brands (Apr 2026) | Single-client case study |
| Adobe Analytics, Q1 2026 | AI referral +42% conversion vs non-AI in March 2026, from −38% a year earlier; +37% revenue per visit; +48% time on site; 393% YoY AI traffic growth | Vendor, very large sample |
| Kaiser & Schulze, Marketing Science | 973 e-commerce sites, ~50,000 ChatGPT transactions vs 164M other transactions; ChatGPT 13% below organic, 32% below paid | Peer-reviewed |
| Shero Commerce, 2026 | 73 stores, 96.4M sessions, 976,668 orders; AI 1.27% vs organic 1.09% at median; ChatGPT 88.1% of AI visits | Large published dataset |
| Lebesgue, April 2026 | 35,000 Shopify brands; AI referral 3.6% vs 1.23% traditional search; close to brand-query conversion at 3.65% | Vendor dataset |
| Zero-click and AI Overview data (Semrush, Pew, Ahrefs, BrightLocal via Searchlab) | Transactional 31% zero-click vs informational 74%; local 72%; commercial investigation 46%; AI Overview presence 16.5% transactional vs 39.4% informational; position-one CTR −58% when an AI Overview appears (Ahrefs, Feb 2026, 300,000 keywords); Pew 8% vs 15% click rate | Multiple independent studies |
| Cheers Market Baseline, 31 Aug–20 Sep 2026 | ≥82% of roofing answers citing a directory; own-site citation 17.8% roofing vs 44.0% HVAC, 53.7% plumbing; page-type split 37.4% homepage / 24.1% service / 17.4% service+city / 10.7% location / 1.2% blog; review medians 91 / 130 / 152; 84.1% at ≥4.8 stars; 0.6% under 4.0 | Published study, vendor |
| BrightLocal Local Consumer Review Survey 2026 | 68% require 4+ stars; 47% avoid under 20 reviews; 74% weight the last three months; 89% expect an owner response | Long-running survey |
| SOCi 2026 Local Visibility Index | Average recommended business 4.3 stars across 350k+ locations | Vendor, large panel |
| Roofing Contractor 2026 surveys | 74% of homeowners follow a recommendation; 50% use search engines; 49% of contractors run Google Ads; 11% use AI search services; 40% use AI in the business | Trade survey |
| WordStream / LocaliQ 2025 series | Home & Home Improvement: $8.33 CPC, 8.05% CVR, $90.92 CPL — US dollars, US accounts, all campaign types | Large published benchmark |
| Ad X Groupe, October 2026 | 3,198 US home-service sites: 22% pass all five AI-readiness checks; 52% lack LocalBusiness markup; 61% have no homepage FAQ | Published industry scan |
| Axios, 28 October 2026 | 41% domain overlap on same-day repeats of identical questions | Reported research |
| Practitioner estimates on dark traffic | 20–40% of AI visits landing as Direct, with some reporting 60–70%; referrer stripping from ad blockers at 42.7% of visitors | No shared methodology — directional only |
Two structural warnings about this evidence base. Almost all of it is US data, which matters for a Canadian roofer because the directory stack and the auction pricing differ. And almost all of it measures conversion rates on sessions, which understates a channel whose primary output is a phone call and a brand search that never appears as a session. The instrumentation in section 09 exists to close both gaps, and until you have 90 days of your own data, the external figures are a hypothesis rather than a benchmark.
17Questions we get asked
How many leads should a roofer expect from AI search in a month?
Small numbers with high value, at least for the next year. AI referrals average around 1.08% of total sessions across industries, so a roofing site doing 4,000 monthly sessions should plan for tens of visits, not hundreds — and one to three leads a month from those, plus the brand searches and directory calls the answer generates. The case for the work is that five of the seven first moves also improve paid search and Google Business Profile, not that it produces volume this quarter.
Does ChatGPT actually send roofing leads?
It names businesses, and some of the people reading those answers call or search the name. In a September 2026 study of 31,470 responses, 12.1% of home-services users clicked the first business named, 7.2% clicked the second, and a further 8.4% searched the first brand's name instead of clicking. The click-through is the visible part; the brand search is usually larger in roofing and shows up as organic brand traffic.
Is AI traffic really worth five times organic traffic?
Not reliably. The 15.9% figure comes from a single B2B client, while Amsive's paired analysis of 54 websites found 4.87% for LLM traffic against 4.60% for organic — a difference that failed statistical testing at p = 0.794. The credible claim is that AI-referred buyers are further along in research-heavy categories like planned replacement and insurance claims. Measure your own segment before you put a multiplier in a forecast.
What is the highest-intent roofing prompt class?
Emergency. The decision is same-day, price sensitivity is at its lowest, and a repair frequently converts into a replacement. The constraint is almost never visibility — it is whether the answer can see your hours and phone number as extractable text, and whether a human answers at 11pm. Second is cost, where publishing a range with a review date beats "call for a free quote" for both citation and qualification.
Why do AI leads show up as Direct traffic in my analytics?
Because AI applications frequently strip the referrer and privacy settings and blockers remove it from a large share of sessions. Practitioner estimates put the share landing in the Direct bucket at 20–40%. That is why call tracking with dynamic number insertion, an intake source field and branded-search tracking are the primary instruments for this channel, with the referral report as a secondary check.
Should I move budget from Google Ads to AI SEO?
No, not as a substitution. Paid search is the only channel that can respond within hours to a storm event, and it is the only way to bid on a competitor's name. AI search compounds slower-moving planned-replacement and insurance-claim demand over quarters. Keep paid for storm weeks and conquesting; fund the entity and conversion work because it also lifts paid performance.
How many reviews does a roofing company need to be named in AI answers?
In the 2026 roofing citation data the medians were 91 reviews for a business named by ChatGPT, 130 for AI Mode and 152 for Gemini, with 84.1% of named businesses at 4.8 stars or above. Consumer expectations run lower but are rising: BrightLocal reports 68% will only use a 4-star business and 47% avoid anyone under 20 reviews. Treat 100 reviews, current within three months, as the working target.
What should I ask an AI SEO agency before hiring them?
Three questions. Which of your numbers is a lead, and how do you know? How will you attribute a call that arrives with no session? And what would make you tell me to stop? If the reporting is visibility scores and citation screenshots without a path to your CRM's installed jobs, you are buying a dashboard rather than a channel.
How long before AI SEO produces a measurable change?
Sixty to ninety days for the instrumented parts — calls becoming visible, brand search moving, directory corrections landing — and a full quarter before you can make a defensible claim about jobs. Same-day repeats of identical questions overlap only about 41%, and Amsive measured 7.53% site-level variance, so anything shorter is a story rather than a measurement.
Does any of this work differently in Canada?
Yes, and mainly in the directory layer. In the 2026 collection, HomeStars appeared in 40% of Canadian ChatGPT roofing answers and none of the US answers, while Angi and Expertise.com — the two leaders in the American set — were effectively absent from Canadian answers. The measurement discipline is identical; the platforms you correct first are not.
18Glossary
The terms in this article, defined the way we use them with contractors.
- Prompt class
- A group of AI queries that produce the same kind of answer and require the same page and conversion path. Roofing has five that matter commercially.
- Position
- The order in which a business is named inside an AI answer. It is the commercial unit of this channel, worth roughly 2.4× between third and first in home services.
- Zero-click rate
- The share of searches or prompts that end without a click to any website. About 31% for transactional queries, 74% for informational ones.
- Dark funnel
- Lead activity that analytics cannot see: stripped referrers, brand searches after reading an answer, and calls placed from a directory card.
- Brand-search follow-up
- The share of users who type the recommended brand into a search engine instead of clicking a citation. 8.4% for the first-named home-services business.
- DNI
- Dynamic number insertion — swapping the phone number displayed on a page according to traffic source, so calls can be attributed.
- CPL
- Cost per lead. Spend divided by leads. Useful for vendor comparison, dangerous as an optimisation target because it ignores whether the lead was qualified.
- Cost per booked job
- Spend divided by jobs installed and attributed. The only denominator that matches a roofing profit and loss statement.
- AJV
- Average job value. The revenue side of the acquisition ceiling calculation.
- RoofingContractor
- The schema.org type for a roofing business beneath HomeAndConstructionBusiness. Carries service area, hours and contact facts in machine-readable form.
- Card list
- The block of business cards an assistant renders beneath a local recommendation. Present in about 96.8% of roofer answers in one 2026 dataset, averaging 11 businesses.
- Cited vs. retrieved
- Being retrieved means a page was pulled into the candidate set; being cited means it was named in the answer. Roughly 15% of retrieved pages are cited, so retrieval alone is not visibility.