AI SEO for Window and Door Companies: How to Get Cited in ChatGPT

01Two indexes, one gate: why Bing decides

The first fact about AI SEO for window and door companies is not about content. It is about which index your pages are in.

ChatGPT's search feature retrieves live web pages, and the published evidence says that retrieval tracks Bing closely. Seer Interactive matched over 500 SearchGPT citations against search results and found 87% aligned with Bing's top results, against 56% for Google. The sample was small and the method observational, and OpenAI does not document its search providers — but the direction has been corroborated often enough that we treat it as the working model. AirOps separately put ChatGPT-to-Bing overlap at 73%.

The operational consequence is blunt. A window company can rank third in Google for "window replacement Markham", have an optimised Google Business Profile, and be completely absent from Bing — because nobody ever verified it, submitted a sitemap, or used IndexNow. In that state, no amount of on-page rewriting changes ChatGPT's behaviour, because the page is never a candidate.

There is a second-order point that matters for a category with as much JavaScript as this one. Bing's crawler is less forgiving of client-rendered content and heavy framework hydration than Google's, and window sites are unusually dependent on both: configurators, colour swatch galleries, multi-step quote forms. The pages that matter commercially are frequently the ones rendered last.

02The four-gate funnel

Every ChatGPT citation passes four gates in sequence. They are dependency-ordered: fixing gate three while failing gate one is the most common way money is wasted on this channel.

Gate four is where content strategy lives, and it is where every agency pitch begins. It is also the gate that matters least if you are failing gate one — which, on window and door sites, is the single most common finding in our audits. If you want the deeper mechanics of gate two and three, they are set out in the citation playbook we wrote for roofers; the funnel is identical across trades, only the source mix changes.

03What the engines actually cite in this trade

Here the window and door industry has better data than most. A 2026 home-services visibility index ran 32 natural-language trade prompts across 100 metro areas in the US and Canada, across four engines, for 28 days ending 2 September 2026, and collected 356,019 citations. Its headline finding is that the engines disagree about what a good source is.

Two rows of that index matter specifically to this trade. Under its trade breakdown, windows and doors (grouped with duct cleaning and lawn care) accounted for 21,340 citations, split 67.6% to business websites and other sources, 26.3% to directories and review platforms, 2.5% to social and forums, and 3.5% to Google's own pages. That is a better split than ChatGPT's home-services average, which tells you own-site content carries more weight in this category than in, say, plumbing.

The same index reports that when ChatGPT did name a tracked brand, the average mention landed at 2.9 in the list. That number should shape your expectations more than any appearance rate: being third on a four-name list is a materially different commercial outcome from being first, and it is measured separately.

04The directory layer, by platform and metro

If roughly a quarter of citations in this trade go to directories, the useful question is which directories — not "should we do listings". The same 2026 index publishes citation counts per platform across its 100 metros, and the Canadian entries are the ones a GTA window company should read twice.

Cheers Market Baseline, 28 days ending 2 September 2026, 356,019 citations. Counts are across all four engines, not ChatGPT alone. Google's own properties, at 11,729 citations, are excluded from this table.

Read the metro column as a coverage map. A platform cited in 100 metros is being reached for in every market, which usually means the engine trusts it structurally. A platform cited in 30 to 46 metros is being reached for in the markets where it has data — which is exactly why a Canadian window company with a strong HomeStars profile can be cited in Toronto and invisible in a US city, and vice versa. The lead generation work we do starts with this: fix the records that are actually being cited, in the metros you actually serve, before touching the homepage.

05Ghost citations: quoted but not named

A citation and a mention are different events, and the gap between them is where most AI SEO reporting goes wrong.

Semrush's analysis of AI answers reported that 61.7% of citations were "ghost" citations — the domain is used as a source but the brand is not named in the answer — and that only 13.2% of domain appearances carried both a citation and a mention. In the same body of work, ChatGPT cited brands in 87% of appearances but mentioned them in only 20.7%, while Gemini showed the reverse pattern at an 83.7% mention rate and a 21.4% citation rate. BrightEdge's tracker separately found ChatGPT mentions brands about 3.2 times more often than it links them.

For a window and door company the ghost problem is expensive in a specific way. A homeowner reading an answer does not click a footnote; she reads the names in the sentence and, at most, taps a card. If your cost page was used to build "expect $8,000 to $18,000 for a whole-home replacement in the GTA", you have paid for the content and received nothing, because your company name was not in the answer.

This is also why the cheapest reporting on the market is misleading. A dashboard showing "cited in 34% of answers" can be true while your name appears in 9% of them, and only the second number produces phone calls. Ask any vendor which one they are measuring.

06The dealer problem: why the manufacturer outranks you

Window and door has a structural handicap no other home-service trade shares: your product has a brand that is not your brand. A homeowner asking about fibreglass windows will get an answer full of manufacturer names. A homeowner asking who to hire will get dealers, installers and remodellers whose only differentiator is proximity and proof.

The data explains the mechanism. Ahrefs' analysis of ChatGPT's 1,000 most-cited pages found 67% sit in categories a local business cannot reach, and among those, brand homepages were 23.8% and Wikipedia 29.7%. Manufacturer domains are structurally large, densely linked and encyclopedic — exactly the profile that gets retrieved and quoted. A dealer's ten-page site is not competing on the same axis.

Three implications, in order of value:

• Compete on the commercial queries, not the product encyclopaedia. "What is a U-factor" will be answered by a manufacturer or a government programme page, and you will not win it. "Vinyl versus fibreglass cost difference in [city], installed" is a query a dealer can own, because the number lives in your price book.
• Make the dealer relationship machine-verifiable. If you claim an authorised-dealer status, it needs to exist on the manufacturer's own locator page and be linked with sameAs. An unverifiable claim of certification is the kind of assertion an assistant has no reason to repeat.
• Disambiguate the entity. A dealer called "Example Windows" in a market with a manufacturer product line called something similar will be merged, confused or dropped. Consistent legal name, address, phone and hours across site, profile and directories is the fix, and it is the same fix local SEO has always required.

This is the point at which generic AI SEO advice stops being useful. A programme for a SaaS company optimises the product page; a programme for a window dealer has to establish a local installer entity inside a category the manufacturers already dominate. Our window and door internet marketing practice is built around that difference, and the strategic version of it is set out in the AI marketing guide for this category.

07What decides selection at the page level

Gate four is where most of the controllable variance sits, and the studies are unusually specific about which page-level properties correlate with being cited rather than merely retrieved.

One finding in that list deserves a warning attached. SE Ranking also reported that 28% of ChatGPT-cited pages had no organic search visibility at all. People quote this as proof that SEO is dead. It is not: it means the citation pool is broader than the Google top ten, so pages that Bing trusts but Google does not rank can still be cited. It does not mean a page nobody can find anywhere gets cited.

08Engineering a quotable passage

A passage gets lifted when it is self-contained, specific and attributable. The test we apply: could a reader who saw only those three sentences, with no surrounding page, correctly answer the question and know who said it?

Compare that with the version most window sites publish: "We offer competitive pricing on premium energy-efficient windows, tailored to your home." There is nothing to extract, nothing to compare and nothing to attribute. It is not that the model dislikes the page; there is simply no fact in it.

The five passage types worth writing, in this order

1. The cost range — per unit and per project, bounded by product type, quantity and city, with a date.
2. The eligibility rule — the exact threshold that decides a rebate or a code question, in one sentence, with the standard named.
3. The comparison verdict — two materials, one axis, a number: lifespan, maintenance interval, cost delta.
4. The process step — what happens in what order, with time units a homeowner can plan around.
5. The limitation — what you do not do, or where a cheaper option is the better buy. This is the passage that gets quoted into "should I hire" answers, because it is the one no competitor writes.

One formatting note that costs nothing: put the passage in prose, not only in a table. Tables get parsed, but a sentence with a subject and a verb is what an answer reuses when it needs to explain rather than list. Publishing both versions of each fact — a table row and a sentence — doubles the formats an engine can lift.

09The five technical blockers on a window site

Window and door websites fail at retrieval for reasons that are specific to how these businesses sell. None of them are content problems.

• Specification PDFs. Energy ratings, warranty terms and installation guides are almost always PDFs. Conversion and extraction from PDFs is unreliable, and a phone at 11pm is the wrong device for one. Every fact in a PDF needs an HTML twin.
• The configurator as the only path to product data. If sizes, colours and glass options exist only inside a JavaScript tool, the machine-readable version of your catalogue is a set of option labels with no values. Publish static range pages per product line.
• Image-only galleries. A portfolio of 400 installations with no captions, no city and no product names carries no extractable information and earns no local relevance. Alt text and captions are the difference between a gallery and a dataset.
• Parameterised duplicates. Colour and size variants that generate ?colour= URLs split authority and confuse canonicalisation. One canonical page per product and city, with variants as content on that page.
• Quote forms as pages. A "get a quote" page that is the only place your pricing lives is a page nobody can cite. Pricing belongs on a public page; the form should confirm and qualify, not disclose.

The five blockers above sit in the technical layer of a window and door digital marketing programme and are cleared before any editorial work begins — there is no point rewriting a cost page that Bing has never indexed.

Two access checks take ten minutes and catch most of the rest. Fetch a priority page with a plain-text client and confirm the pricing paragraph is in the response; then fetch /robots.txt and confirm OAI-SearchBot is not disallowed.

# 1. Is the answer in the raw HTML, or only after JavaScript runs?
curl -s https://yourdomain.ca/window-replacement-cost-markham \
  | grep -o 'U-factor[^<]*' | head -3

# 2. Can the retrieval crawler reach you at all?
curl -s https://yourdomain.ca/robots.txt | grep -iA2 'OAI-SearchBot'

# 3. Is the page in Bing's index? (expect the URL itself, not zero results)
curl -s -A "Mozilla/5.0" "https://www.bing.com/search?q=site%3Ayourdomain.ca+window+replacement" | grep -c 'yourdomain.ca'

10Specification markup that survives extraction

Schema is not a citation lever — a controlled test of 1,885 pages adding JSON-LD against roughly 4,000 matched controls measured −4.6% in AI Overviews, +2.4% in AI Mode and +2.2% on ChatGPT. What schema does is remove ambiguity about what a number is. On a specification-heavy product like a window, that is worth having.

The important technical detail for this trade: schema.org has no window or door installer type. Companies invent one, which is worse than using a general type. The workable pattern:

{
  "@context": "https://schema.org",
  "@type": "Product",
  "name": "Triple-pane fibreglass casement window, Low-E argon",
  "brand": { "@type": "Brand", "name": "<Manufacturer>" },
  "additionalProperty": [
    {
      "@type": "PropertyValue",
      "name": "U-factor",
      "value": 1.14,
      "unitCode": "UN/CEFACT:W/M2K",
      "measurementTechnique": "NAFS / CSA A440 certified test"
    },
    {
      "@type": "PropertyValue",
      "name": "Energy Rating",
      "value": 38,
      "unitCode": "C62"
    },
    {
      "@type": "PropertyValue",
      "name": "Solar Heat Gain Coefficient",
      "value": 0.29
    }
  ],
  "isRelatedTo": {
    "@type": "Service",
    "name": "Whole-home window replacement",
    "areaServed": [
      { "@type": "City", "name": "Markham" },
      { "@type": "City", "name": "Stouffville" }
    ]
  }
}

Three rules for the implementation. Attribute per unit, not per company — a U-factor belongs to a specific unit with a specific glazing package, and stating it as a business-level property is a claim you cannot defend. Use the business type you actually are — HomeAndConstructionBusiness with areaServed as a city list, not an invented subtype. Keep dateModified honest, because it is the cheapest freshness signal you control and the easiest to fake, and faked freshness is exactly the kind of inconsistency an assistant resolves against you.

The rest of the pattern — rebate pages, warranty promises, product-by-city pages — is set out in the category marketing guide; this section is only the extraction layer.

11The page set, mapped to prompts

Not every page deserves the engineering. Five page types produce citations that convert in this category, and they map to the prompts ChatGPT actually receives.

The mapping is the strategy — the same logic we applied when mapping window company marketing strategies onto prompt classes. A company that writes ten blog posts about "signs you need new windows" is working the one class where no company gets named. A company that publishes one accurate rebate page and one cost page per product line is working the two classes where being named is normal.

12The authority layer: links, mentions, records

Three things outside your site decide whether the model has any reason to trust it.

Records. Yext's study of 6.8 million citations found 86% came from sources a brand owns or manages, split between first-party websites (44%) and business listings (42%). In home services specifically, ChatGPT's directory share is high enough that an outdated hours field is a genuine citation loss. The fix is a quarterly field-level audit across the platforms in section 04 — name, address, phone, hours, service area, services, licensed status — not a subscription that marks profiles as "verified".

Mentions. Branded web mentions were the strongest correlate of AI visibility in Ahrefs' data at 0.664. For a window company the reachable version of this is unglamorous: local media coverage of an installation, supplier and manufacturer dealer pages, trade association directories, chamber listings, and the handful of regional "best of" lists that actually get cited. One correction: paid listicle placements often produce no citation, because the engines increasingly discount self-promotional roundups.

Reviews. BrightLocal's 2026 research puts 68% of consumers requiring four stars or higher, 47% avoiding businesses under 20 reviews, 74% weighting only the last three months, and 89% expecting an owner response. For citation purposes the recency threshold matters most: a business with 140 reviews and none since March looks dormant to a system that favours fresh corroboration. Ask at handover, ask for a description of the work, and reply to every review.

13Measurement that connects to leads

Three numbers describe AI visibility, and only the third produces revenue:

• Appearance rate — share of tracked prompts where you are named at all. The 2026 index reported a mean of 25.1% for ChatGPT and a median of 21.6% across 119 home-services brands.
• Citation rate — share where your domain is the linked source, at a mean of 13.4% in the same dataset. Lower than appearance, always.
• Position — where in the list, at an average of 2.9 when named. This is the number that predicts calls, and almost nobody tracks it.

Then connect it to money with three instruments that survive referrer stripping and six-week decision cycles:

# The weekly panel: 20 prompts, 5 classes, fixed order, logged — not screenshotted
week | class | prompt | named? | position | my domain cited? | competitor named

Call tracking with dynamic number insertion, so a call from a directory card or a brand search is attributable at all. An intake source field — "How did you first hear about us?", with an AI assistant option — because brand follow-up after reading an answer is a documented behaviour and it lands in analytics as ordinary organic. And Bing rank as the leading indicator: because ChatGPT retrieves through Bing, a fall in Bing position is a leading signal that ChatGPT visibility is about to follow, typically within a week or three rather than in a quarter.

For the full instrumentation, including the appointment and quote stages that decide whether any of this was worth doing, see the high-intent lead measurement guide. For the lead-volume side specifically, the whole-home replacement leads analysis covers what a citation is actually worth at the quote stage. The instruments are trade-agnostic; only the page set changes.

14The 90-day technical build

This is the order we run inside digital marketing for window and door companies engagements, and the order matters more than the pace. Weeks one and two are the ones a content-led agency skips.

15What does not work, with the numbers

• llms.txt as a visibility play. In Ahrefs' analysis of 137,210 domains, 97% of published llms.txt files received zero requests, and most of the traffic that did arrive came from SEO audit tools. Ship it for agent readiness if you like; do not expect it to move citations. Full evidence in the llms.txt guide.
• Schema as a lever. −4.6% / +2.4% / +2.2% across three surfaces in the controlled JSON-LD test. Implement it for extraction clarity, not ranking.
• Content volume. More articles do not help if each one answers a question where no company is named. In adjacent trade data, contractor-site citations concentrated on homepages and service pages rather than blog posts.
• Google-only optimisation. If your entire measurement apparatus is Search Console, you are blind to the index ChatGPT actually retrieves from. Bing Webmaster Tools is not optional in this channel.
• Citation guarantees. The same question asked twice on the same day produces only about 41% domain overlap, which is why no one can promise a fixed position inside an answer.
• Reporting that stops at appearance. Appearance without position, and position without a call-tracking join, is a dashboard rather than a channel.

16Sources and confidence

Every figure above traces to one of the following. Where a source is a company with a commercial interest in the conclusion, it is labelled.

Three caveats travel with this evidence base. The Bing dependency is a measured pattern from a small observational sample, not a documented architecture, and OpenAI does not publish its retrieval providers. Most citation datasets are US-weighted, which is why the Canadian directory rows in section 04 matter. And almost none of it is specific to windows and doors — the trade appears only as a grouping row, so the category-specific claims here are inferences from a broader home-services base rather than a dedicated study.

17Questions we get asked

18Glossary

OAI-SearchBot

OpenAI's crawler for ChatGPT search. Separate from GPTBot (training) and ChatGPT-User (user-triggered fetch). Opting out removes you from ChatGPT search answers.

IndexNow

A protocol for notifying Bing and other participating engines when a URL is created or changed, shortening the gap between publishing and indexation.

Candidate set

The pages retrieved for a prompt before selection. Roughly 16.6 URLs per prompt in Ahrefs' data; only a fraction are cited.

Ghost citation

A source used in an answer whose brand is not named inside it. Reported at 61.7% of citations in one Semrush analysis.

Sub-query

A rewritten internal query the model issues to search for supporting evidence. Title similarity to these sub-queries correlates with citation.

Appearance rate

Share of tracked prompts where a brand is named at all, regardless of citation. Mean 25.1% for ChatGPT in 2026 home-services data.

Citation rate

Share of tracked prompts where the domain is the linked source. Mean 13.4% in the same dataset.

Position

Place in the list of named businesses. Average 2.9 in the same dataset; the number that best predicts calls.

U-factor

Whole-unit heat loss rate. Canada's ENERGY STAR threshold is 1.22 W/m²K; Most Efficient is 1.05.

additionalProperty

The schema.org mechanism for attaching measurable specs — U-factor, Energy Rating, SHGC — to a Product.

HomeAndConstructionBusiness

The schema.org type to use for a window and door company, since no installer-specific subtype exists.

Prompt panel

A fixed set of prompts asked in the same order weekly, producing a log of appearance, citation and position instead of screenshots.