The 60-second answer
Become the cheapest business in your market to verify. An assistant names first whichever company it can confirm — that it exists, does this specific work, in this specific place, at a stated standard — from the most independent sources with the fewest contradictions. So make every record agree, put a recent review profile behind the claim, narrow the category until you are the obvious answer to one question rather than a plausible answer to eight, and get written about by name in prose that no directory can generate.
What does not work: publishing more, buying listicle placements, adding markup, or trying to out-cite a competitor. In one 2026 study the brands that lost the top slot had more citations than the brands that won it. The competitive asset is not volume of evidence — it is the absence of ambiguity around a specific claim.
Read those four together and the strategy changes shape. The first slot is worth roughly double the second and more than the rest combined, so it is worth pursuing specifically rather than hoping for. It is not bought with more evidence, because the losers had more. It is not inherited from Google rank, because the two are almost decoupled. And most of what decides it — 86% of the citation base — sits on assets you already own or can edit this month.
What the first slot is worth
ChatGPT has no results page, so "position" needs defining before it can be valued. In the study below it means the order in which brands appear when an answer recommends more than one — the first name in the paragraph, then the second, and so on.
| Position in the answer | Direct click-through | Branded search follow-up | Total visit rate |
|---|---|---|---|
| 1st recommended | 10.3% | 12.4% | 22.7% |
| 2nd recommended | 5.7% | 6.9% | 12.6% |
| 3rd recommended | 4.9% | 5.8% | 10.7% |
First Page Sage, October 2026: 31,470 ChatGPT responses from an opt-in panel of 2,150 US adults, combined with referral data from 112 websites across 10 industries. Vendor research, self-reported panel, US-weighted — directional rather than definitive, and not roofing-specific.
Three things in that table matter more than the headline. The first brand's 22.7% exceeds the total for positions five through ten combined, so the tail of a long answer is worth very little. Branded search follow-up is larger than direct clicking at every position — more than half the value arrives later, through a search engine, wearing the appearance of organic brand demand rather than AI referral. And the effect concentrates further when the answer is short: when three brands were recommended, the first got a direct click in 12.9% of answers, against 7.4% when eight or more were listed.
Visit rate by position in the answer
The branded-search finding deserves its own sentence, because it changes where a contractor should spend. If most of the value of a first-position recommendation arrives as a branded Google search, then the recommendation and your classic search presence are not competing activities — the recommendation creates the demand and your organic brand result and Business Profile convert it. A roofing company with a strong first-position recommendation and a neglected brand search profile loses most of the benefit at the last step — which is the connection between being named and turning high-intent roofing leads into booked work.
Why the most-cited brand loses
The intuitive strategy is to accumulate evidence: more coverage, more listings, more mentions, more sources that name you. It is the strategy the industry sells, because it is the one that produces a report with a rising line. And it does not appear to win first position.
Leoprd's Reputation to Revenue 2026 study, published in September, compared brands that regularly converted visibility into a recommendation with brands that appeared frequently and were rarely chosen. The losers averaged 9.2 citations. The winners averaged 8.4. Across unbranded recommendation prompts, monitored brands were the first choice 29% of the time, mentioned but passed over 35%, and absent 36%. Both groups drew on a remarkably similar mix of commercial, editorial and review sources — commercial content accounted for 46% of citations among brands with strong preference and 46.8% among brands that were visible but rarely chosen.
If the winners and losers have the same source mix and the losers have slightly more of it, then citation count is not the mechanism. Something about how the same sources describe the business decides preference — most plausibly clarity, specificity and consistency, which are properties of the description rather than its volume. That is why the rest of this article is about what the sources say, not how many there are.
Leoprd's vertical breakdown is a useful caution about expectations. In travel and transport, brands were present in 90% of relevant answers but chosen first only 28.1% of the time; in fintech, 92.3% presence and 32.1% first choice; in B2B software, 67.8% first choice when present; in energy and utilities, 92.9%. Presence and preference are separately earned, and the conversion rate between them varies enormously by category. Roofing, being local, transactional and trust-dependent, should be expected to sit toward the lower end — which is exactly why the specific local signals in sections nine and ten carry the weight. The prior problem, getting named at all, is the subject of the recommendation piece; this article starts where that one finishes.
Why Google rank does not buy it
The second comfortable assumption to discard is that ranking first produces being named first. A June 2026 analysis of more than 1,300 ChatGPT recommendations across 10 industries and 10 states found the two close to decoupled. Businesses in a Google top ten were recommended about 30% of the time. Recommended businesses ranked somewhere in a top fifty 20% of the time, and in a top ten only 5%. And within the ranked set, position barely mattered: the top three were recommended about 26% of the time against 24% for positions four to ten, with first place (31%) no better than eighth (30%).
That study also found Google and Bing rankings overlapping by only about 4%, which is a useful sanity check on the whole concept — there is no single "ranking" for a local business to win, which is consistent with the model not choosing from either results page.
| Finding | Figure | What it rules out |
|---|---|---|
| Google top-ten businesses that were recommended | 30% | That ranking is sufficient |
| Recommended businesses ranking in any top fifty | 20% | That ranking is necessary |
| Recommended businesses ranking in a top ten | 5% | That the mechanism is a SERP read |
| Recommendation rate, ranked first versus eighth | 31% versus 30% | That position within a SERP is an input |
| Google and Bing ranking overlap for the same businesses | ≈4% | That "ranking" is a single, stable quantity |
Omnipresence, The Credibility Layer, June 2026: 1,300+ ChatGPT recommendations, 10 industries, 10 states, 858 distinct sources cited. Vendor research with an unusual and useful negative result; the sample is US and cross-industry.
Two practical consequences follow. Ranking work is not wasted — it is a precondition for being cited, which is a different and sometimes sufficient outcome, and in Google's own AI Mode the company's own site supplied 87.2% of home-services citations. But a contractor who reaches position one on Google and stops has not reached position one in ChatGPT, and should not expect the same programme to deliver both. The better reading of the evidence is that ranking and recommendation share an upstream cause — prominence — rather than one producing the other.
Which points at the question the rest of the article answers: if not volume, and not rank, then what? The pipeline that produces the answer, from fetch to assembly, is set out in the retrieval engineering brief.
The model optimises for verification cost
The most useful frame available is an inference rather than a documented mechanism, and it is worth labelling as such. An assistant recommending a contractor is making a claim on someone's behalf about a business it cannot inspect. Every recommendation carries a risk of being wrong in a way the user will notice — wrong city, wrong trade, wrong licence, a company that closed. So the system prefers the candidate whose claims it can confirm at the lowest cost: the fewest sources to open, the fewest contradictions to resolve, the most recent evidence that the claim still holds.
That single sentence explains almost every finding in this article. It explains why consistency dominates — a contradiction is a verification cost. It explains why recency matters — a two-year-old review says less about today. It explains why categories must be narrow — "roofing contractor in Markham" is cheaper to verify than "general contractor serving the GTA". It explains why the most-cited brand loses: twenty sources that disagree cost more than eight that agree.
One supporting number makes the frame less speculative than it sounds. Yext's analysis of 6.8 million AI citations found that 86% came from sources the business controls or can influence: its own website at 44% and its business listings at 42%, with reviews and social at 8% and Reddit and forums at 2% once location and intent were accounted for. If the overwhelming majority of the evidence base is yours to edit, then verification cost is largely a maintenance discipline rather than a PR campaign. Before spending anything on the programme, confirm that the business is reachable and parsable at all, which is what the seven-test diagnostic establishes.
Signal one: consistency across records
This is the highest-leverage and least glamorous work available, and the practitioner evidence is unusually direct — not a study of AI outputs but a survey of the people who work on local visibility. In Whitespark's 2026 Local Search Ranking Factors survey, 47 local search experts ranked the signals driving AI visibility, and three of the top five were citation factors: name, address, phone and description matching across the web.
The mechanism is the verification-cost frame in its purest form. "Acme Roofing Ltd., 100 Main St, 905-000-0000, roofing contractor, Markham" repeated identically in twelve places is one fact, confirmed twelve times. The same business with a suite number here, a tracking number there, a trading name in one bio and a legal name in another is four businesses with partial files, and the fourth is a competitor with a cleaner record.
| Field | Where to check | The failure that costs you the slot |
|---|---|---|
| Legal name | Site footer, Business Profile, licence registry, invoices, BBB, directories | A trading name in prose that matches no registry entry |
| Address | Business Profile, site, directories, manufacturer locator | A suite number present in three places and absent in five |
| Telephone | Everywhere a listing exists | Call-tracking numbers on listings, splitting one business into several |
| Primary category | Business Profile primary and secondary | "General contractor" set as primary when the work is roofing |
| Service area | Site copy, areaServed markup, Business Profile service area | Three different lists, one including places you have never worked |
| Hours | Profile, site, directories | Emergency availability claimed in one place, contradicted in another |
| Description | Profile summary, directory bios, social bios | Three different one-line descriptions of the same company |
| Mentions and pages | What the co-citation graph implies about other companies | No page anywhere describes the other companies you are listed beside |
The practical programme is a single spreadsheet with one row per external record and one column per field, filled by hand, then a diff. It takes about three hours for a typical contractor and it is the least interesting work in this article. It is also the work that most reliably precedes a change in which businesses share an answer, because it is the only signal that is fully under your control and fully verifiable by the system reading it. Where the same records are also being rewritten for search generally, the sequencing we use is documented in the AI SEO case study.
There is research on co-citation worth knowing about: an August 2025 study found 82% of AI Mode citations came from top-ranked pages that were never the business's own site, and work on "mentions and pages" suggests part of the mechanism is which other companies appear alongside you. That evidence is thinner and more contested than the consistency findings, so treat co-citation as a hypothesis to test in your market rather than a lever to buy.
Signal two: the review profile
Reviews are the most misunderstood signal in this category, because the industry treats volume as the goal and the evidence treats it as a threshold. Three findings frame what actually matters.
First, the rating of recommended businesses clusters high but not perfect. In SOCi's 2026 study of more than 350,000 business locations, the locations ChatGPT recommended averaged 4.3 stars — and ChatGPT recommended just 1.2% of them, against 35.9% for Google's local three-pack. Second, volume matters up to a point: in a roofing-specific panel, named companies had a median of 130 Google reviews, with 23.6% of answers naming a business with fewer than twenty, while only 0.6% of rated appearances showed anything below 4.0 stars. Third, recency is decisive for buyers: BrightLocal's 2026 survey found 74% of respondents weighted reviews from the last three months.
| Belief | Evidence | What to do instead |
|---|---|---|
| "More reviews than competitors" | Median 130 among named roofers, with a large minority below twenty | Get safely past the local median, then stop optimising for count |
| "A perfect rating wins" | Recommended locations averaged 4.3, not 5.0; 99.4% of rated roofing appearances were at or above 4.0 | Treat the rating as a floor. Clear it comfortably and defend it |
| "Old reviews are fine" | 74% of consumers weight the last three months | Ask every completed job, every month, forever |
| "Replies are a nicety" | 89% of consumers expect owners to respond | Reply to all of them, in a sentence, within a week |
| "One platform is enough" | Review sources are platform-specific: on Bing Places data, Facebook led, then Yelp, TripAdvisor, Porch, Yellow Pages and Angi | Match platforms to the engines your buyers use, rather than defaulting to Google alone |
SOCi 2026, 350,000+ locations. Cheers Market Baseline, 2,176 answers, Sep 2026. BrightLocal 2026 consumer survey. Whitespark review-source analysis for Bing Places. All vendor or survey research; none peer-reviewed.
The practical reading is a two-part rule. Clear the floor — rating safely above 4.0, review count at or above the local median — and then put the effort into velocity and recency, because those are the properties that make a business cheap to verify today. A company with 400 reviews averaging 4.6, the newest from fourteen months ago, reads as uncertain; a company with 140 averaging 4.7 with eleven in the last six weeks reads as current. Review evidence is also the mechanism behind being cited at all in some engine mixes, which the citing playbook covers in detail.
Signal three: category and positioning precision
A recommendation is a match between a described need and a described business. Vague descriptions on either side make the match expensive, and the system resolves that cost by choosing a business that describes itself precisely.
Roofing companies resist precision for understandable commercial reasons: they do everything, they have always done everything, and narrowing feels like turning away work. But the recommendation layer is not a sales meeting. It is a matching problem, and "we do all types of roofing and general contracting across the GTA" is a description that matches nothing in particular while competing with every generalist in the province.
| Description | Questions it can win | Why |
|---|---|---|
| "Roofing and general contractor serving the Greater Toronto Area" | Almost none outright | Matches a hundred businesses; no specific need resolves to it |
| "Roofing contractor in Markham" | "Who does roofs in Markham" | Category plus place; the minimum viable match |
| "Sloped-roof replacement and storm-damage repair in Markham and Stouffville, including insurance claim documentation" | Replacement, storm damage, insurance claims, in two named places | Each clause is a sub-question the business can own outright |
Three mechanics make the third description win. It is specific enough to match generated sub-questions, which fan out into comparisons, costs and situations rather than head terms. It is narrow enough that the business is the obvious answer rather than a plausible one. And it is stable enough to be corroborated, which matters because precision only compounds if the same narrow claim is repeated elsewhere — in the Business Profile primary category, in manufacturer programme listings, in the first line of third-party mentions.
The disciplined way to choose the narrow claim is to pick the one question where you already have the strongest proof: the most recent reviews mentioning it, the most documented jobs, the most credible third-party corroboration. Then own that question completely before widening. A roofing company that is first choice for "insurance claim roof repair in [city]" and fourth for everything else is in a far stronger commercial position than one that is fourth for eight things. Choosing which narrow claim to own is easier with the wider tactic inventory in the 50-strategy set beside it.
Signal four: corroboration density
Corroboration is what a directory cannot give you: prose, written by someone with no commercial interest, that says what you do and names you. Its importance is measurable. Ahrefs' analysis of 1.4 million prompts put the correlation between branded mentions and AI visibility at 0.664 — the strongest single measured relationship in this field. BrightEdge found ChatGPT mentions brands 3.2 times more often than it links to them.
That last ratio is the one to internalise. The system talks about companies it has read about far more often than it links to them. A link is plumbing; a mention is testimony. And testimony is what turns a citation into a recommendation.
| Source class | What it contributes | Realistic effort |
|---|---|---|
| Manufacturer and supplier case studies | A named contractor described doing a specific job with a specific system — the single most quotable form of third-party prose in this trade | One per quarter, via the rep who wants the story |
| Trade association and programme directories | Independent confirmation of credential and locale, in the vocabulary buyers search | Annual; mostly admin |
| Local editorial coverage | A news or community outlet naming the business in prose, with a date and a reason | One or two a year, via a real local story |
| Podcasts and trade media | Long-form mentions in a transcript that becomes indexable text | One a quarter if there is something to say |
| Community and forum answers | Recommendations in local groups and neighbourhood forums, from real accounts | Ongoing, and must be genuine |
| Supplier and partner pages | A named listing on a vendor's or builder's site, in context | Low, and often overlooked |
Two disciplines keep this honest and effective. Named, in prose, with a reason — "Acme Roofing replaced the shingle roof on this 1960s bungalow using [system] in two days" is corroboration; a logo strip is decoration. And one per month, sustained, beats six in a fortnight and nothing for a year, because recency is part of what makes evidence cheap to verify.
Signal five: hyper-local specificity
The strongest structural finding in the recent research is that this is not one national competition but a hundred local ones. Across 858 distinct sources cited in the Omnipresence study, 95% appeared in only one industry and 68% appeared only once. The winners were local: "the audit resolves locally", and while national brands occasionally win broadly, most verticals have purely local champions.
For a contractor this is the most encouraging finding available, and it has a precise implication: the unit of competition is not "roofing" but "roofing in this municipality", and the evidence that counts is evidence that resolves at that grain. A beautifully written page about roof replacement in general, on a site with three vague city pages, is competing in the wrong race. The on-page version of the same point, including which city pages are worth keeping, is in the on-page optimization guide.
What local specificity actually looks like, in ascending order of usefulness:
- The municipality named in the page title, the H1 and the answer block — not just in a list of service areas.
- The permit office that issues the permit, named, because it is a locally specific string nobody else publishes.
- The housing stock and what it implies: interwar semi with three-tab shingles versus 1990s subdivision with architectural shingles versus a flat-roofed commercial strip.
- Local weather reality: snow load, ice damming, freeze-thaw cycles, hail frequency — stated in the vocabulary an adjuster uses.
- Named neighbourhoods and the specific problems they have, from your own job records.
- Something that happened: a storm date, a street, a project with a duration and a material system.
The warning that comes with this, from earlier in this cluster, still applies: forty pages differing only by municipality are a doorway pattern and they blur the entity. The distinction is whether the page contains facts that are true only of that place. If it does, it is a real page in a real race. If it does not, it is a template competing with the model's existing supply of templates.
How to earn mentions, with the actual asks
This section is the operational core, so it is written as requests you can send rather than principles you can admire. Each of the four below produces a named, dated, third-party sentence — the unit that correlates 0.664 with being visible at all.
The supplier case study
To the manufacturer's territory rep: "You need a completed-installation story for the new [system]. We have a [city] job from [month] with [squares], [pitch], a [detail] we solved, and photos. Your team writes it, we do not edit it, and you get the asset." Suppliers need content and rarely have good field material; a contractor who brings both gets named on a manufacturer page that these systems read.
The association listing upgrade
To the trade association: "We are a member in good standing, licence [number]. Can the member directory list our specialisms and service area rather than the default one-liner?" The listing already exists; the ask is precision, because a directory entry with the right category and place is worth more than a generic entry with a logo.
The local story
To a community outlet: "We replaced [n] roofs on [street] after [event] and noticed [specific local finding — a code detail, an insurance pattern, a material failure]. Here is the data, no cost, no advertorial." Editors want the finding, not the business; being named in the sentence is the reward for supplying it.
The partner page
To a builder, property manager or restoration firm: "Add us to your recommended contractors page with two sentences on what we handle and where." This is the cheapest mention available and the most commonly ignored, because nobody thinks to ask for it.
Two rules on execution. Never pay for a mention and present it as independent — paid placements inside "best roofers in [city]" roundups are increasingly discounted by both readers and models, and disclosed sponsorship is a different and weaker signal. And keep a register: URL, publisher, date, the exact sentence, and the last time it was checked. That register is the asset, not the mention itself, because it tells you what to re-earn when a page goes dark.
How to build the review engine ethically
Reviews are the one signal where velocity can be engineered without misleading anyone, and where the failure mode is not a policy breach but a quiet decline. The mechanics are simple enough to be boring, which is why they are rarely done.
- Ask at the moment of relief, not the moment of invoice. For a roof replacement that is the day after the crew leaves and the house is dry; for a repair it is when the leak stops. Asking during final payment produces compliance, not enthusiasm.
- Ask every customer, every job, through one channel. A single text with one link, sent by the person who ran the job, in their own voice. Multiple asks across platforms fragment your rating profile.
- Reply to all of them within a week. 89% of consumers expect it, and a reply is itself text a system can read — it restates the service, the place and the outcome.
- Never filter, never incentivise. Gating reviews by expected sentiment violates platform policy and destroys the credibility of the only proof layer you have.
- Put the number on the site, in text, with a date. "4.7 from 148 Google reviews, last checked 1 November 2026" is extractable; a widget that loads after paint is not.
- Add the second platform deliberately. If your buyers use Bing or Apple surfaces, the review sources differ — on Bing Places data, Facebook led, then Yelp and others. One platform is a single point of failure.
The target is not a perfect rating. It is a defensible, recent, corroborated one: comfortably above 4.0, at or above the local median in count, with movement every month and a reply against everything. That profile is cheap to verify, which is the whole argument.
Programme and credential listings
Roofing has a structural advantage most trades lack: certification programmes that the answer engines actually read. In a five-engine panel of 2,176 answers, 56.3% of Gemini's roofing answers involved a manufacturer programme. That is a distribution channel disguised as a credential.
Three moves turn it into position. Enrol in the programmes that exist in your material mix and are plausible for your volume — being a certified installer of the systems you actually install, at the tier you can honestly claim. Then make the listing precise: programme name, tier, service area, and a link that resolves, because a locator entry pointing at a dead URL is worse than none. Then mirror the same strings on your own site, in the sentence where they do work, so the lexical layer and the entity layer agree.
| Credential | What a system can verify | Priority |
|---|---|---|
| Manufacturer installer certification | Programme name, tier, and a locator listing — the one credential that appears inside roof-specific answers | Highest |
| Licence and registration | Issuing body and number, checkable against a registry | Highest, and non-negotiable |
| Insurance and workers' coverage | Stated limits, verifiable on request | High — buyers ask directly |
| Trade association membership | Membership status and directory entry | Medium |
| Manufacturer warranty authorisation | Ability to register extended coverage for the homeowner | Medium, and commercially useful |
| BBB accreditation | Accreditation status and rating, cited heavily by Perplexity | Medium, engine-specific |
| Generic "trust badges" | Nothing verifiable | None |
One caution about the last two rows in combination. Perplexity leans on BBB more than other engines — 59.6% of its roofing answers in one panel — so in a market where Perplexity matters, a clean BBB file is worth the annual effort. In a market where ChatGPT dominates, the manufacturer locator likely matters more. Engine mix decides priority, which is why section fifteen asks you to log position per engine rather than in aggregate.
When the top slot is already taken
It often is. The same cross-industry study found that within a single market the leading recommendation holds 78% consensus across differently worded prompts — the leader is locked in, and phrasings do not dislodge it. Expecting to displace a five-year incumbent with a better website is the most common strategic error in this category.
The realistic plays, in order of speed:
| Play | How it works | Time to first position |
|---|---|---|
| Own a sub-question | Become first choice for "insurance claim roof repair" or "metal roof replacement" while the leader holds the generic query | One to two quarters |
| Own an adjacent municipality | The leader's evidence resolves to one city; build genuine local proof in the next one | Two quarters |
| Own a segment | Commercial flat roofs, heritage homes, property managers, new-build warranty work | One to two quarters |
| Displace on recency | Wait for the leader's evidence to age — reviews stall, programmes lapse, sites decay — and out-maintain them | Two to four quarters |
The flank is often the better business anyway. Being first choice for insurance-claim roof repair in one city can be worth more per lead than being third choice for "roofers near me", because the query arrives with a decision already made and a claim already open. And a flank, once held, is a platform: the same corroboration machinery that won the narrow question generalises to the next one, whereas a head-on assault on a locked leader usually produces a year of reports and no movement.
How to tell which situation you are in, before committing a quarter: run the panel with three differently worded versions of the same question. If the same business is named first each time, you are facing a locked leader and should flank. If the first name varies with phrasing, the position is contested and head-on work is viable.
How to measure position, not presence
Almost every AI-visibility report measures presence, because presence is easy to count. Position takes work and tells you something. Three changes to a standard panel convert it into a position instrument.
- Record the ordinal, not the boolean. For each prompt, log the exact order of names, not whether you appeared. A run where you are third in every answer looks identical to a run where you are first in every answer if you only count appearances.
- Log the brand mention separately from the citation. A cited domain with an unnamed brand is a ghost citation, and the aggregate rate in one 2026 analysis was 61.7%. If your position score and your mention score diverge, the problem is attribution rather than selection.
- Track branded follow-up, not just clicks. Since branded search follow-up exceeds direct clicks at every position, the measurement that reflects the value is whether AI exposure lifts branded queries — visible in Search Console as a branded-impression change that lands days after the panel moves.
Two bounds keep this honest. Sixty prompts per engine is a practical floor, and at a 45% hit rate that yields a 95% interval of roughly 32% to 58% for presence; position adds a second layer of noise, because identical questions repeated the same day overlap only about 41% in the domains cited. So report position as a median with a range — "median position 2 in 34 of 60 prompts, range 1 to 4" — rather than as a percentage, and observe a control cohort so that change can be attributed to work rather than to the index moving under you.
A single table: prompt, engine, date, ordered list of named businesses, whether you were cited, whether you were named. Everything else in an AI-visibility dashboard is a derivative of that table, and most of it is less useful than the table itself. It takes twenty minutes a week and it is the only way to know whether the work in this article is doing anything.
The 90-day ladder to first choice
The sequence is deliberate: cheap-to-verify work first, because it changes which businesses share the answer, and corroboration last, because it compounds slowly and cannot be measured until the base is clean.
| Weeks | Work | Done when |
|---|---|---|
| 1 | Baseline the panel with three phrasings per question, logging order; check whether the leader is locked | You know whether to compete head-on or flank |
| 2–3 | Consistency audit across every external record; fix name, address, phone scheme, category, service area, description | One row per record, zero contradictions |
| 4 | Choose the narrow claim: the one question your proof is strongest for | A single sentence that names trade, work type and places |
| 5–6 | Rewrite the homepage answer block and the chosen service page around that claim, with the company named in the fact-bearing sentence | The claim appears verbatim in three places on your own site |
| 7 | Review engine: one ask per completed job, replies inside a week, the figure published in text with a date | New reviews in each of the last four weeks |
| 8 | Credential pass: programme tier confirmed, locator entry correct and resolving, association listing precise | Every credential is verifiable from a third-party page |
| 9–10 | Send the four asks: supplier case study, association listing, local story, partner page | Two named mentions committed in writing |
| 11 | Local specificity: permit office, housing stock, storm reality, named neighbourhoods on the city page | Three facts that are true only of that place |
| 12 | Re-run the panel; freeze the result; change nothing for a month | A dated before-and-after with the control cohort beside it |
Two notes on pacing. The first-position change, when it comes, usually arrives as a swap in which three or four prompts move at once rather than a gradual creep, because the underlying evidence changes as a block — which is why weekly measurement matters more than the average. And nothing before week eight should be expected to move the first slot at all: weeks two to seven are buying the right to compete.
What does not move you up
Publishing more pages
More pages produce more citations, and more citations did not choose the winner in the study above. Volume competes on the axis where you are already losing; specificity competes where you can win.
Buying listicle placements
Paid entries in "best roofers in [city]" roundups are discounted by readers and models alike, carry disclosure obligations, and produce a placement rather than a recommendation. The exception is a genuinely independent ranking you happened to qualify for — which is earned, not bought.
Schema as a lever
Adding JSON-LD to 1,885 pages against roughly 4,000 controls produced no significant citation movement. Markup removes ambiguity about who you are, which is worth doing once and worth nothing as a growth tactic.
Chasing the Google number one
Ranking first versus eighth barely changed recommendation odds in the cross-industry data, and only 5% of recommended businesses ranked in a top ten at all. Rank is worth having for citations; it is not the route to first choice.
Fast review bursts
A hundred reviews in a fortnight from a link in a van is a spike, not a profile, and platforms treat it accordingly. Monthly velocity with replies is both better evidence and safer.
Any guarantee of first position
The leader's consensus is 78% within a market and the shortlist regenerates per prompt. Nobody can guarantee an ordinal they cannot observe, and any contract that promises one is describing a mechanism that has not been disclosed.
Sources and confidence
| Figure used | Source, date, sample | Confidence |
|---|---|---|
| Visit rate by position: 22.7% first, 12.6% second, 10.7% third; branded follow-up exceeds direct clicks at every position; 12.9% versus 7.4% direct click with three versus eight brands listed | First Page Sage, Oct 2026; 31,470 ChatGPT responses, panel of 2,150 US adults, referral data from 112 sites | Vendor panel |
| First-choice rate 29%; mentioned but passed over 35%; absent 36%; passed-over brands averaged 9.2 citations against 8.4 for the first choice; source mix near-identical | Leoprd, Reputation to Revenue 2026, Sep 2026 | Vendor study |
| Top-ten ranked recommended 30% of the time; recommended businesses in a top ten 5%; first versus eighth 31% versus 30%; Google/Bing overlap ≈4%; 95% of 858 cited sources appear in one industry | Omnipresence, The Credibility Layer, Jun 2026; 1,300+ recommendations, 10 industries, 10 states | Vendor research |
| Recommended locations averaged 4.3 stars; ChatGPT recommended 1.2% of 350,000+ locations versus 35.9% for the Google three-pack | SOCi, 2026 | Vendor study |
| 86% of 6.8M AI citations from controllable sources: own site 44%, listings 42%, reviews and social 8%, forums 2% | Yext, 2026 | Vendor study |
| Three of the top five AI-visibility signals are citation factors, per 47 practitioners | Whitespark, 2026 Local Search Ranking Factors | Practitioner survey |
| Branded mentions correlate 0.664 with AI visibility across 1.4M prompts; ChatGPT mentions brands 3.2× more often than it links them | Ahrefs; BrightEdge | Vendor |
| Median 130 Google reviews among named roofers; 23.6% of answers named a business with under 20; 99.4% of rated appearances at or above 4.0 | Cheers Market Baseline, 2,176 answers, Sep 2026 | Vendor panel |
| 74% weight reviews from the last three months; 89% expect owner replies; 45% used AI for local recommendations | BrightLocal, 2026 consumer survey | Survey |
| 56.3% of Gemini roofing answers involved a manufacturer programme; BBB cited in 59.6% of Perplexity answers | Cheers panel, Sep 2026; Local Vitals Florida, Apr 2026 | Vendor panels |
| 78% consensus for the leading local recommendation across phrasings | Omnipresence, Jun 2026 | Vendor research |
| 61.7% ghost citations; ChatGPT named a brand in 20.7% of appearances | Semrush, 2026 | Vendor |
| Adding JSON-LD to 1,885 pages produced no significant citation movement | Ahrefs controlled study, May 2026 | Controlled |
| ≈41% same-day domain overlap for repeated identical questions | Axios, Oct 2026 | Journalism |
Two limits apply to everything above. None of it is roofing-specific except the panels that say so, and none of it is peer-reviewed in the sense a technical reader expects — these are vendor studies, surveys and practitioner research, several of them with a product to sell. The verification-cost frame in section five is an inference drawn across those results, not a disclosed mechanism. Where the evidence is thin, this article says so rather than building on it: the co-citation material in section six is the clearest example, and the 78% consensus figure is one study's finding rather than a law.
Questions we get asked
How can you make your roofing company one of ChatGPT's top recommendations?
Be the cheapest business in your market to verify. Make the legal name, address, phone scheme, category and service area identical in every place a machine looks; narrow the claim until you are the obvious answer to one question; keep the rating comfortably above 4.0 with reviews arriving every month and a reply on each; and earn named mentions in third-party prose. Position is won by removing ambiguity around a specific claim, not by adding evidence.
Does being the most-cited brand make you the first recommendation?
Apparently not. In a September 2026 study, brands that appeared in AI answers but lost to a competitor averaged 9.2 citations, while brands chosen first averaged 8.4. Both groups drew on a nearly identical mix of commercial, editorial and review sources. The difference is in how the sources describe the business rather than how many of them there are.
How much more is the first recommendation worth than the third?
Roughly twice. In one October 2026 study of 31,470 ChatGPT answers, the first-recommended brand received a visit in 22.7% of answers against 12.6% for second and 10.7% for third, and the first brand's total exceeded positions five through ten combined. More than half of that value arrived as a branded search afterwards rather than a direct click.
If you rank first on Google, will ChatGPT recommend you first?
No, and the two are close to decoupled. Cross-industry research found Google top-ten businesses were recommended about 30% of the time, recommended businesses ranked in a top ten only 5% of the time, and recommendation rates for ranking first versus eighth were 31% versus 30%. Ranking is worth having for citations; it is not the route to being chosen first.
What is the single highest-leverage fix?
Consistency across records. In a 2026 survey of 47 local search practitioners, three of the top five signals for AI visibility were citation factors — name, address, phone and description matching across the web. It is also the cheapest work available, because it is entirely under your control and takes about three hours with a spreadsheet.
How many reviews does a roofing company need?
Enough to clear the local median rather than a national target: named roofers in one 2026 panel had a median of 130 Google reviews, and 23.6% of answers named a business with fewer than twenty. Above that threshold, velocity and recency matter more than count — 74% of consumers weight reviews from the last three months, and recommended locations averaged 4.3 stars rather than a perfect score.
How do you earn third-party mentions without paying for them?
Four asks, each of which produces a named, dated sentence: a supplier case study from a manufacturer rep who needs field content, a precision upgrade to an existing association listing, a local story with a genuine finding for a community outlet, and a two-sentence entry on a builder's or property manager's recommended-contractor page. Keep a register of URL, publisher, date and exact sentence.
What if a competitor already holds the top slot in your city?
Flank rather than charge. One cross-industry study found the leading local recommendation holds about 78% consensus across differently worded prompts, so displacement by a better website is unlikely. Win a sub-question such as insurance-claim repair, an adjacent municipality, or a segment such as commercial flat roofs, then generalise from the position you hold. Test three phrasings first: a stable winner means a locked leader.
How do you measure position rather than presence?
Log the ordered list of names per prompt per engine, not a yes-or-no appearance; record brand mentions separately from citations, because ghost citations run above 60% in one analysis; and watch branded search impressions, since branded follow-up exceeds direct clicks at every position. Report a median position with a range across at least sixty prompts, and keep an untreated cohort.
Can anyone guarantee a first-position recommendation?
No. The shortlist is regenerated per prompt, the leader in a market may hold 78% consensus or the position may be contested, and identical questions repeated the same day overlap only about 41% in the domains cited. What can be verified and reported is the work: record consistency, review velocity and rating, credential accuracy, named mentions earned, and median position with its range.
Glossary
- Position
- The ordinal place of a business in the list of names an AI answer recommends, as distinct from whether it appears at all.
- First choice
- The business named first in an answer that recommends more than one; in one 2026 study the first brand received a visit in 22.7% of answers.
- Passed over
- A brand that appeared in an answer but was not selected as the recommendation — 35% of unbranded prompts in one study.
- Verification cost
- The effort a system expends to confirm a business's claims. Used here as a prioritisation model rather than a disclosed ranking function.
- Consistency factor
- Agreement of name, address, phone and description across external records; three of the top five AI-visibility signals in a 2026 practitioner survey.
- Corroboration density
- The number and quality of independent third-party statements about the business, each naming it in prose with a reason.
- Ghost citation
- A citation in which the domain is used as a source while the brand is never named in the answer.
- Branded follow-up
- A subsequent search for the brand's name after seeing a recommendation; larger than direct clicking at every measured position.
- Locked leader
- A market where the same business is named first regardless of prompt phrasing — about 78% consensus in one cross-industry study.
- Flank
- Winning first position for a sub-question, adjacent municipality or segment rather than contesting a locked leader head-on.
- Precision claim
- A single sentence naming trade, work type and place, narrow enough that the business is the obvious match rather than a plausible one.
- Review floor
- The rating threshold treated as a minimum rather than a goal; recommended locations averaged 4.3 stars in one 2026 study.
- Control cohort
- A set of comparable prompts or pages left untreated so that movement can be attributed to work rather than to index drift.