Quick answer: Traditional search gave a homeowner ten links and let them choose. AI gives them one answer, and they ask again, and again — a repair usually takes six to ten questions before anyone picks up a phone. Each of those questions is a separate chance to be the source that gets cited. So the job is no longer ranking a page per service and per city. It is covering the question surface: the Learn, Compare and Act questions a homeowner actually asks, for each service line, in each market you serve. That is a few hundred answers, not a few dozen pages.
The shape of the search changed, not just the ranking
Old behavior: one query, ten blue links, the homeowner does the comparing. Your job was to be one of the ten.
New behavior: a question, an answer, a follow-up. "Why is my AC freezing up." Then "can I fix it myself." Then "is it worth repairing a 12-year-old unit." Then "what does that cost." Then "who in Annapolis does this." Then "are they any good."
Six questions. Six answers. Six different sets of sources cited. The homeowner never saw a page of ten links, and never compared vendors the old way — the assistant did the comparing, using whatever it could cite at each step.
That is the change that matters. Ranking was a contest you entered once per query. Citation is a contest you enter once per question — and there are far more questions than there were queries.
What we measured
On Aug 3–4, 2026 we put the same problem to ChatGPT, Gemini and Perplexity two ways.
Asked nationally — "Why is my air conditioner freezing up and how do I fix it?" — the assistants drew on 56 citations (Gemini) and 17 (Perplexity). Every source was a national publisher, manufacturer, or big directory. No local contractor appeared anywhere.
Asked locally — "My AC is freezing up and I live in Annapolis, Maryland — who should I call?" — the citation count collapsed to 18, 20 and 6. And the answers named local companies: MaxAir, Pro-Tech HVAC, B&B Air Conditioning, Aire Serv of Annapolis, One Hour Heating & Air, Coastal Heating & Air.
Two conclusions worth acting on:
- Generic national questions are not winnable for a local contractor. You are competing with manufacturers and national media for a slot among 56 sources. Do not spend your budget there.
- The same question, localized, is winnable. The citation pool shrinks by two thirds and the competitive set becomes companies your size.
The most useful thing we found
Read how Perplexity justified one of its local recommendations:
"B&B Air Conditioning & Heating Service — offers 24/7 emergency AC repair in Annapolis and says frozen coils are a common cause of AC problems."
That second clause is the whole strategy in one sentence. The assistant did not cite B&B because of a directory listing or a review count. It cited them because B&B had published content about frozen coils — and when a homeowner asked about frozen coils, B&B was a source that could be quoted on it.
They earned a local recommendation with an educational page. Not with a services page. Not with a city page. With an answer.
Why "a page per service and city" stopped being coverage
The standard build is services × cities. Twelve services, eight cities, ninety-six pages, and it feels thorough.
But look at what those pages answer. "AC repair in Annapolis" answers exactly one question — who does AC repair here. That is the last question in the chain, and often the only one the old model needed you to win.
The five questions before it — what is wrong, can I fix it, repair or replace, what will it cost, how do I choose — are the ones the homeowner asks first, and they are asked far more often. A services-and-cities site is silent on nearly all of them. It shows up for the final question and is absent from the whole decision that led there.
That is not a ranking problem. It is a coverage problem, and no amount of optimizing ninety-six pages fixes it.
The three levels a real answer library needs
We organize educational content on an intent ladder, because homeowner questions fall into three distinct kinds and each needs a different page.
- Learn — how it works, basics, the education a homeowner searches before they call. Why the coil freezes. What a heat pump actually does. Why a system short-cycles. This is where the volume is, and where B&B won.
- Compare — decision-stage questions. Repair or replace. Heat pump versus furnace. Brand against brand. Two-stage versus variable-speed. These are the questions where an assistant is actively looking for someone who has laid out the tradeoffs plainly.
- Act — cost guides, financing, second opinions, maintenance plans. What a replacement runs in this market. What financing looks like on it. This is where a cited answer converts, because the homeowner has already decided to do something.
Now multiply. Say forty real questions per service line across those three levels, four service lines, and the local variants that matter for your markets. You are not writing ninety-six pages. You are building several hundred answers — and each one is a separate opportunity to be the cited source.
That is why we ship Learning Center content in packs rather than one-offs. Coverage is the product. A single excellent page about frozen coils wins one question; a library wins the chain.
"AI search" is not one place
It is easy to hear "AI search" and picture ChatGPT. The surfaces behave differently and you need to think about them separately:
- Google AI Overviews — sits above the organic results and often answers without a click. Its presence is per-query, not per-site: in our SERP data an AI Overview appeared on "best hvac marketing companies" and did not appear on "hvac marketing agency." Same topic, different surface, so what "winning" looks like changes query by query.
- Google AI Mode — a full conversational surface where the follow-up chain happens inside Google.
- The Knowledge Panel and local pack — entity-level, not page-level. This is fed by your Business Profile, your structured data, and consistency of your facts across the web. It answers "who are you" rather than "how does this work."
- ChatGPT, Perplexity, Gemini, Copilot — each with its own citation behavior. In our tests Gemini cited the most sources by a wide margin; ChatGPT sometimes cited almost nothing and answered from its own summary.
The through-line: page-level content earns question citations, entity-level signals earn "who should I call" recommendations, and you need both. A great answer library with an inconsistent Business Profile gets cited on the how-to and skipped on the hire.
Where lists and directories actually fit
They matter, and they are one layer — the vendor-selection layer.
When we asked who the best HVAC company in Frisco, Texas was, Perplexity's cited sources were two Reddit threads from the city's subreddit, an Angi list page and a Thumbtack list page. Not one contractor's own website. At that specific question — who should I hire — third-party lists and community threads carry the answer, so being present in them is worth real effort.
But that is one question out of the chain, and it is the last one. Treating "get on the lists" as an AI strategy is like treating a phone book entry as a marketing plan. Necessary. Nowhere near sufficient.
One caution, since it is tempting: do not manufacture your own recommendations in community threads. Those communities detect it, and a thread about a contractor caught doing it becomes a permanent, well-indexed, highly citable document pointed the wrong way.
Being cited is only half of it
A citation sends someone who is mid-decision, already holding a specific question, to a page. If that page is a wall of prose ending in "contact us," the citation was a courtesy to your competitor who converts better.
The answer page has to carry the next step inside it — the estimate tool, the financing calculator, the second-opinion offer, the booking path — matched to the intent level that brought them. A Learn visitor is not ready to book; they are ready to understand and then compare. An Act visitor is ready to book and should not have to hunt for how.
This is why we treat educational content and conversion components as one system rather than two projects. Coverage gets you cited. Components decide whether the citation was worth anything.
What we cannot tell you yet
- Citation is not instant. Still's HVAC went live on Hydra the day before these tests and was not cited in the Annapolis answers. New content has to be published, crawled and established. Anyone promising fast AI citations is guessing.
- These were two questions on two days in specific markets. Your citation sets will differ by city and trade. Run the tests for your own market rather than trusting ours.
- Assistants change their behavior frequently. A result from August 2026 is a snapshot, not a rule.
- Nobody can honestly attribute revenue to AI citations today with the precision we attribute paid search. What is measurable is whether you are cited — and that is worth tracking now, because it is the input.
What to do on Monday
- Write down the real question chain for your most profitable job. Not keywords — the actual sentences a homeowner says, from first symptom to hiring. Usually six to ten.
- Ask an assistant each one, once nationally and once with your city attached. Note whether you are cited and which sources are. That list is your competitive set, and it is more useful than any rank report.
- Count your coverage. How many of those questions does your site actually answer on its own page? For most contractor sites the honest answer is one — the last one.
- Fill the Learn gap first. It is the largest, the cheapest to write, and the layer where a contractor demonstrably wins a local citation.
- Fix the entity layer in parallel — Business Profile, structured data, consistent facts. That is what earns the "who should I call" answer once your content earns the "what is wrong" one.
The old game was being one of ten links for one query. The new one is being the best available answer to as many of your customer's questions as you can afford to answer well. Ten pages of services and cities was never going to cover that.
If you want help mapping the question surface for your market and building the library that covers it, that is the work — see our AI search guides, how generative search reads a site, content intelligence, or the sites we have launched.
