OpSpring.ai
All case studiesSolar / Home Services

East Coast Solar

A Worcester solar installer went from about one lead a month to 18 estimate requests in its first 58 days, on a site rebuilt for Google and for the AI assistants buyers now research with.

18
estimate requests in 58 days
~9x
lead rate vs the old site
18,708
search impressions in 58 days
5x
faster mobile load, service + portfolio

The challenge

East Coast Solar designs, engineers, permits and installs solar across Massachusetts, all in-house rather than subcontracted. In a market full of sales outfits that hand the actual work to whoever is cheapest that month, that is a genuine differentiator. Almost none of it reached anybody searching.

The Squarespace site produced roughly one estimate request a month, and its estimate form was worse than that number suggests: it simulated a submission and discarded the lead. Hero imagery shipped as 971KB CSS background images, bypassing any optimizer. Page titles rendered the brand twice. There was no analytics property, so none of this was visible.

The timing made it urgent. The federal 30% residential tax credit, worth roughly $9,000 on a $30,000 system, was eliminated at the end of 2025, and passive, incentive-driven demand went with it. The buyers still shopping are researching, and research happens in search results and increasingly inside AI assistants. A company invisible in both has no pipeline.

There was a structural problem underneath. The domain reads eastcoast.solar, implying coverage from Maine to Florida, while the business is Worcester and Central Massachusetts. And their Google Business Profile is a hidden-address service-area listing that does not surface in the Places API by exact name, while Worcester competitors do. Every geographic signal the site could send needed to point somewhere the profile itself was not pointing.

The solution

We replaced the Squarespace site with a custom Next.js build: about 40 route groups expanding into roughly 70 pages, including 35 town service-area pages, residential and commercial service trees, a portfolio, an incentives library and a savings calculator.

The calculator is the piece that changed the shape of the funnel. A visitor enters an address and a monthly bill and gets back a modeled system size, annual production, 25-year savings and CO2 offset, computed against Google's Solar API. It gives people a reason to hand over an address that has nothing to do with wanting a salesperson to call, the opposite of how a contact form works. Results are not email-gated; the client rejected that outright and asked for a "speak to a professional" prompt beside the numbers instead, which was the right call.

The incentives content was built to be the most useful page on the Massachusetts solar internet: SMART, net metering, the state credit, the federal ITC's actual status, MACRS depreciation and Class I RECs, each explained in a self-contained passage rather than scattered across a sales narrative.

Then the part that took the longest. The first version of the 35 town pages shared about half their text, because only a description and a short local paragraph varied between them. Google's doorway-page handling exists to suppress exactly that pattern, so every one of them was rebuilt from real local research.

That meant actually researching thirty-five towns. Each page now carries housing stock an installer would recognize (Framingham's Campanelli ranches have wide, low-pitched roof planes and are among the easiest installs in MetroWest; Saxonville is Victorians with steep, cut-up roofs), the real permitting path (which department, the OpenGov portal, whether a historic district adopted under MGL Chapter 40C puts panel placement in front of a commission first), utility specifics, municipal programs with the date each figure was confirmed, and town-specific questions with real answers. Every factual claim traces to a URL in that town's sources array, and the citations render on the page.

Across all 528 page pairs, 7-gram text uniqueness went from 50.7% to 78.3%. The worst pair went from 57.0% to 26.8%, and no pair finished above 40%. Median page length went from 725 words to 1,374.

AI architecture

The site was built on the assumption that a growing share of its readers would be machines summarizing it for somebody else, and that those machines need different things than Googlebot does.

FAQ answers render in the server HTML. The accordions on every service page are force-mounted and collapsed with CSS rather than unmounted, so the answer text is in the document a crawler receives instead of appearing only after a click no crawler performs. Each set is backed by FAQPage JSON-LD. An answer an assistant cannot read is an answer that cannot be quoted.

llms.txt is generated from the same constants and data modules the pages render from, not maintained by hand. It had been a static file, and it drifted: it advertised "15+ reviews" while the LocalBusiness schema on every page said 14. Assistants read both, so a disagreement between them is a credibility problem on exactly the surface the file exists to serve. Deriving it from shared data means the two can no longer disagree.

The structured data was made honest rather than flattering. The self-serving aggregateRating was removed from LocalBusiness. Review nodes on /reviews reference the organization by @id instead of declaring a second LocalBusiness entity, so the page resolves to one entity with one rating. Review counts were aligned to what Google publicly shows.

Accuracy got enforced at the data layer, because an assistant that quotes a wrong claim propagates it. The town template had promised SMART eligibility to every town; SMART is limited to investor-owned utilities, which made it wrong for Shrewsbury, Concord and Hudson, all served by municipal light plants. It now branches on utility. Federal ITC status lives in one place so no two pages can describe it differently. The rule written into the data files: omit a field rather than guess at it, and date any figure that can change.

AI referral traffic is tracked as its own channel in GA4. That is the only reason there is a number to report at all: most sites cannot tell an assistant-referred visit from direct traffic, so the channel is invisible to them by default. IndexNow submits new and changed pages on deploy rather than waiting to be crawled.

Performance was treated as part of this. The h1 animated in from opacity zero, so the server-rendered headline, usually the LCP element, was invisible until JavaScript hydrated. Hero backgrounds shipped as unoptimized CSS images. Both were fixed: hero images now go through the optimizer at roughly 50KB of AVIF instead of 971KB of JPEG, and the headline animates from a stylesheet keyframe that runs at parse. Throttled mobile LCP on /residential went from 9,152ms to 1,828ms, /portfolio from 7,840ms to 1,340ms, and the homepage from 7.3s to 4.2s with FCP from 2.9s to 0.9s.

The results

In the first 58 days after launch, the site produced 18 estimate requests from 18 distinct users. The site it replaced produced roughly one a month. That is about nine times the lead rate, and 4.0% of visitors now ask for an estimate.

People also read more once they arrive. Pages per visit rose from 1.67 to 1.93, and sessions were up about 7% on the same window a year earlier. All of this happened in the first season after the federal 30% credit ended, so the lift came from the site itself rather than from demand coming back.

In those same 58 days the new site earned 18,708 search impressions across 967 distinct queries, spread over 50 pages. The town pages that got the research work are leading. MetroWest averages position 8.6, on page one, with Greater Boston (11.9) and Framingham (16.5) next in line.

AI assistants are now a channel East Coast Solar can see. GA4 reports assistant-referred visits as their own source, and those visits are already arriving. A competitor without the same setup gets the same kind of visit and files it under direct traffic.

Engineering notes

The decisions that shaped this build, and the reasoning behind each. Most of the interesting ones are about restraint: what was deliberately not shipped, and which numbers were thrown out before reporting.

01
Custom Next.js rather than another hosted site builder
The whole engagement is a bet on technical ceiling. Server-rendered FAQ answers, generated llms.txt, per-page JSON-LD, security headers, AVIF, IndexNow and route-level control of the document are not reachable on a template platform. Squarespace was not underperforming because it was configured badly.
02
35 researched town pages, not 351 generated ones
Massachusetts has 351 municipalities, and generating a page for each was trivially possible. It would also have been an index-bloat machine. The pages ranking closest to the top are all town pages carrying real local research, led by MetroWest on page one, which is the argument against the cheap version.
03
Uniqueness measured with 7-gram overlap across all 528 pairs
The doorway-page risk is about text similarity, so it needed a similarity number rather than an impression that the pages felt different now. Pairwise 7-gram overlap gave a threshold to work against and a worst case to fix, rather than an average that hides the two pages that are nearly identical.
04
FAQ accordions force-mounted and hidden with CSS
Collapsed content that is unmounted from the DOM does not exist to a crawler or an assistant. Rendering it and hiding it visually costs a little HTML weight and keeps every answer quotable.
05
llms.txt generated from the data the pages already use
The hand-maintained version drifted out of agreement with the schema on the same site. Any AI-facing file maintained separately from the content it describes will do this eventually. Generating it makes the drift structurally impossible rather than something to remember to check.
06
Self-serving aggregateRating removed from LocalBusiness
A rating a site asserts about itself, unattached to a review platform, gets discounted at best and treated as manipulation at worst. Review counts were aligned to what Google publicly shows instead.
07
SMART eligibility branched on utility rather than promised everywhere
The town template had asserted SMART for all 35 towns. It applies only to investor-owned utilities, so it was wrong for Shrewsbury, Concord and Hudson. On pages built to be quoted by assistants, a confidently wrong local claim is worse than a missing one, because it travels.
08
Legacy HTTP impressions excluded from every reported figure
About 2,000 impressions sat at average position 1 with zero clicks, all on the old pre-HTTPS address: Google consolidating the previous URL, not anything the new site earned. Including them would have inflated total impressions by 10.8% and taught us nothing.
The cleanest overhaul I could have asked for.
Matt·Co-owner, East Coast Solar

Tech stack

  • Next.js 15
  • TypeScript
  • Tailwind CSS
  • Vercel
  • Google Solar API
  • Google Places API
  • Resend
  • GA4
  • Search Console
  • IndexNow
  • JSON-LD structured data

Ready to streamline your operations?

Schedule a 15-minute diagnosis call to see if AI can solve your biggest bottlenecks.