A client rang me last October sounding genuinely rattled. Their organic traffic was down 40% year-on-year, their phones were quieter, and their newest competitor, a two-van outfit barely 18 months old, was apparently being recommended by ChatGPT when homeowners asked about furnace repairs in their service area. That phone call kicked off the most useful three months of consulting work I’ve done in years, because it forced me to admit something I’d been resisting: the rules of business discovery have quietly changed, and most of us are still optimising for the old game.
What follows is the actual walkthrough of that engagement (with identifying details composited and softened), the decisions I made, the ones I’d make differently, and the playbook I now run for any service business worried about disappearing from AI-mediated search.
The client that forced this realization
A regional HVAC company losing leads
The business: a 22-year-old HVAC company operating across three counties in the American Midwest, roughly $4.8M in annual revenue, 14 trucks, decent reputation. Their website wasn’t broken. Their Google Business Profile was tidy. They ranked in the top three for “[city] HVAC repair,” the exact terms their marketing agency had been chasing for a decade. By every traditional SEO metric, things looked fine.
But their booked-job count had dropped from roughly 280 per month to 168 over eleven months. Repeat customers and word-of-mouth were holding steady; the bleed was entirely in new customer acquisition.
Tracking the 40% organic traffic drop
The first thing I did, and this is unglamorous detective work, was pull GA4 sessions, Search Console impressions, and call tracking data into a single sheet covering 18 months. The pattern was specific:
- Branded search: roughly flat
- “Near me” queries: down 12%
- Long-tail informational queries (“why is my furnace short cycling,” “average cost to replace heat exchanger”): down 61%
- Direct traffic: up 8%
The collapse was concentrated in exactly the kind of queries that people now ask ChatGPT, Perplexity, or Google’s AI Overviews instead of typing into a search box. The traffic hadn’t migrated to a competitor’s website. It had migrated out of websites entirely.
First clue: ChatGPT cited a competitor
I asked ChatGPT (GPT-4o at the time) a plain question: “Who are the most reputable HVAC companies in [their largest city] for emergency furnace repair?” It returned five names. My client wasn’t among them. The two-van competitor was, along with three established firms and one outfit I’d never heard of.
I asked it where the information came from. It mentioned Yelp, the Better Business Bureau, an industry-specific directory I’d genuinely never heard of (more on that in a minute), and a regional “best of” listicle from a local paper. Notably absent: Google Business Profile reviews, my client’s own website, and most of the directories their previous agency had been paying to maintain.
Did you know? According to test-and-learn approach beats analysis paralysis, companies benefit more from quickly establishing a baseline AI discovery strategy and adopting a test-and-learn approach than from spending long stretches on analysis. The market is moving faster than traditional audit cycles allow.
Auditing what AI actually sees
Running queries across four LLMs
I built a query matrix of 47 prompts a real homeowner might use, ranging from “best HVAC company near [city]” to “is [client name] a good HVAC company” to “who should I call for a furnace that won’t ignite in [zip code].” I ran each prompt through ChatGPT, Claude, Gemini, and Perplexity, then logged every business named and every source cited.
This took about six hours. It is not glamorous and there is no tool that does it properly yet (I’ve tried several options, and a handful of others; they’re getting there but I still verify manually). The output was a spreadsheet of roughly 1,400 cells.
Mapping which directories got cited
This is where it got interesting. Across 188 model responses, citations clustered like this:
| Source Type | Citation Frequency | Avg. Position in Response | My Client Listed? |
|---|---|---|---|
| Yelp | 71% | 1st or 2nd | Yes (incomplete) |
| BBB | 54% | 2nd or 3rd | Yes (no recent reviews) |
| Industry-specific directory | 38% | 1st-3rd | No |
| Local news “best of” lists | 29% | Variable | No |
| General business directories | 22% | 3rd-5th | Partial |
Their own website? Cited in 11% of responses, almost always as a secondary verification source rather than the basis for a recommendation. The agency had spent five years building backlinks and writing blog posts. The LLMs treated all of it as background noise.
The Yelp, BBB, and industry-specific surprise
Two things genuinely surprised me. First, BBB still matters enormously to LLMs even though most marketers had written it off as a 1990s relic. The accreditation status, the complaint resolution history, the years-in-business figure all get pulled into model responses. Second, the industry-specific directory I hadn’t heard of (a trade association listing) was being cited more often per query than Angi or HomeAdvisor combined. It had maybe 4,000 listings nationally. It was old, plain, and apparently very well-structured.
Myth: Directory listings are dead because Google killed them in the 2012 Penguin update. Reality: Google’s link-spam algorithms killed low-quality directories as backlink sources. They didn’t kill directories as structured data sources, and that’s exactly what LLMs need to answer business questions confidently.
Choosing where to place bets
Why we ignored 60% of directory options
The previous agency had my client listed on 73 directories. I went through every one and asked three questions: Does any LLM cite this? Does it appear in any AI Overview? Does it pass referral traffic worth tracking? If all three answers were no, I marked it for abandonment.
Forty-four directories failed all three tests. Most were generic “submit your business” sites that nobody, human or machine, actually consults. We stopped paying for the seven that charged subscription fees (saving about $180/month) and ignored the rest. Some of them will stay live for years; that’s fine.
Scoring criteria: citation frequency vs. cost
For the directories that survived the cull, plus a shortlist of new ones to add, I built a simple scoring model:
| Directory | LLM Citation Score (0-10) | Annual Cost | Effort to Maintain | Decision |
|---|---|---|---|---|
| Yelp (claimed + enhanced) | 9 | $4,200 | Medium | Invest heavily |
| BBB Accreditation | 8 | $650 | Low | Renew + rebuild reviews |
| Trade-specific directory | 8 | $240 | Low | Add immediately |
| Angi | 5 | $2,400+ leads | High | Reduce spend |
| Curated business directories | 4 | $0-$99 | Low | Selective adds |
For the curated tier, I included submissions to a handful of editorially reviewed directories, with Jasmine Business Directory being one I’ve used consistently because the listings are human-vetted and the structured data is clean enough that LLM crawlers actually parse it. The cost is negligible; the upside is that these listings often appear in long-tail citation chains where the bigger directories don’t reach.
The $400/month allocation breakdown
The client agreed to a $400/month directory budget (down from their previous $740/month, which is the part they liked). Allocation:
- Yelp Enhanced Profile: $350/month
- Trade directory premium listing: $20/month equivalent
- Curated directory submissions (annualised): $15/month
- BBB renewal (annualised): $15/month
Everything else, Angi, HomeAdvisor, the long tail of submission sites, got cut or zeroed out. The previous agency had been spreading butter over too much toast.
Quick tip: Before paying for any directory, run five test queries through ChatGPT and Perplexity asking about businesses in that directory’s category and region. If the directory isn’t cited even once, your money is buying a 2010-era backlink, not 2025-era visibility.
Rewriting listings for machine readers
Structured data over marketing copy
This was the part the client found genuinely uncomfortable. Their old listings read like brochures: “Family-owned and operated since 2002, we pride ourselves on exceptional customer service and craftsmanship you can trust.” Lovely sentiment. Tells an LLM essentially nothing.
I rewrote every active listing to lead with extractable facts: years in business (23), licence numbers, service area (specific counties named), brands serviced (Carrier, Trane, Lennox, Goodman, Rheem, American Standard, listed explicitly), emergency response time (“under 90 minutes within [county] during business hours”), warranty terms, financing options, employee count, NATE certifications held.
The marketing copy didn’t disappear. It moved to the bottom of each listing, where humans who’d already decided to call could find it.
Service taxonomies that match query patterns
One of the more useful exercises: I pulled the actual phrases people use when asking LLMs about HVAC issues, then made sure those exact phrases appeared in the service descriptions. Not “HVAC repair” but “furnace short cycling diagnosis,” “AC not blowing cold air,” “heat pump making clicking noise,” “ductwork inspection for older homes.” This sounds like keyword stuffing from 2008, but it isn’t. LLMs match user intent to listing content semantically, and dense, accurate service taxonomies give them more handholds.
Specific numbers, ranges, and qualifiers
Vague claims get filtered out. Specific ones get cited. “Affordable pricing” is invisible; “$89 diagnostic fee, waived with repair” is quotable. “Fast service” is invisible; “average response time 73 minutes for emergency calls” is quotable. I rewrote roughly 40 listings with this principle, and the difference in citation behaviour over the next 60 days was the single most measurable outcome of the entire project.
Myth: AI models can’t be trusted to surface specific pricing or service details, so businesses should keep listings vague to maintain flexibility. Reality: Models will surface something with or without your input. Vague listings get replaced by competitor specifics. Posted ranges with “as of [date]” qualifiers protect you better than ambiguity.
Did you know? Roughly 40% of CEOs don’t believe their companies will be economically viable a decade from now if they continue on their current path. Discovery channel disruption is a meaningful slice of why.
Results after 90 days
Citation appearances jumped from 2 to 17
I re-ran the same 47-prompt audit at the 30-day, 60-day, and 90-day marks. Citation count for my client across the four LLMs:
| Audit Point | ChatGPT | Claude | Gemini | Perplexity | Total |
|---|---|---|---|---|---|
| Baseline | 0 | 1 | 0 | 1 | 2 |
| Day 30 | 1 | 2 | 1 | 3 | 7 |
| Day 60 | 3 | 3 | 2 | 4 | 12 |
| Day 90 | 5 | 4 | 3 | 5 | 17 |
| Day 90 (top-3 placement) | 3 | 3 | 2 | 4 | 12 |
Phone calls attributable to “found you online but not via Google” climbed by 31% over the same window. Booked jobs rose from 168/month to 214/month. Not back to 280, but the trajectory was right.
Lead quality shift we didn’t expect
This was the surprise. Leads coming through the AI-cited path closed at a higher rate (52% versus their historical 38%) and had higher average ticket sizes ($1,840 versus $1,210). My working theory: when an LLM recommends a business, the prospect arrives pre-qualified. They’ve effectively been “pre-sold” by an authoritative-feeling intermediary, the way they used to be by a friend’s recommendation. They call ready to book, not ready to comparison-shop on price.
I want to be careful here. This is a sample of 90 days at one company. I’ve seen similar patterns at two other clients since, but I wouldn’t bet a strategy on it being universal until I have a year of data across more verticals.
What still isn’t working
Three things haven’t moved despite our efforts:
- Google’s AI Overviews still don’t cite the client consistently. I think this is a different problem requiring different fixes (more on Google’s idiosyncrasies in another piece).
- Voice assistants (Alexa, Siri) remain a black box. They appear to use entirely different ranking signals and we haven’t cracked them.
- Hyper-local queries at the suburb level. The client appears for the main city but not for the smaller towns within their service area. We’re working on it.
Myth: If you rank well in Google, you’ll automatically be cited by AI assistants. Reality: The overlap is partial and shrinking. LLMs increasingly weight structured directory data, review aggregators, and editorial mentions over traditional organic search signals. A Google #1 ranking and zero LLM citations is now a common (and dangerous) combination.
Adjusting the playbook for different scenarios
Solo operator with $50/month
The HVAC client had a real budget. Most businesses don’t. If you’re a solo plumber, therapist, or accountant with $50/month to spend on visibility, here’s how I’d reallocate:
- BBB accreditation (annualised): ~$15/month
- One trade-specific or curated directory: ~$10/month
- Yelp claimed (free) + active review solicitation: $0
- Remaining $25: Reserve for one editorial placement quarterly (local paper guest piece, podcast appearance, association newsletter)
The free moves matter more at this budget than the paid ones. Claim every relevant free listing, fill them out completely with structured data, and solicit reviews systematically. A solo operator who does this beats a $400/month competitor who treats listings as fire-and-forget.
B2B SaaS where directories matter less
For B2B SaaS, the playbook inverts. Yelp and BBB are largely irrelevant. What LLMs cite for software queries:
- G2 and Capterra (heavily)
- Product Hunt (for newer tools)
- Reddit threads (controversially)
- Comparison articles on tech publications
- The vendor’s own documentation and pricing pages (when well-structured)
I’d put zero budget into business directories and roughly 70% of visibility spend into G2 review acquisition, Reddit-friendly content, and detailed comparison pages on owned media. The principle transfers; the venues change.
Tight timeline: the 30-day compressed version
If a client says “I need results in a month, not three,” the audit phase compresses to two days, the listing rewrites happen in week one, and the focus narrows to whichever two directories scored highest in the citation audit. You won’t see the full 8x citation lift in 30 days; a realistic expectation is 2-3x. Anyone promising more is selling you something.
What if… your industry doesn’t have an obvious trade directory and Yelp is irrelevant (think specialised B2B services, niche manufacturing, professional consulting)? Run the same audit anyway. The citations will cluster around different sources: LinkedIn company pages, association member lists, conference speaker bios, perhaps Wikipedia. The methodology holds; the targets shift. I once ran this for a maritime logistics firm and the dominant citation source turned out to be a 1990s-era industry handbook that had been digitised and ingested into training data.
Principles I’m taking forward
Directories as training data, not traffic sources
The mental shift that took me longest: a directory listing today isn’t primarily a way to send referral traffic to your website. It’s a way to feed structured, verified information into the systems that increasingly mediate every business decision a customer makes. The traffic-source question is essentially obsolete for most directories. The training-data question is the one that matters.
This reframes the whole evaluation. I no longer ask “how many visitors will this directory send me?” I ask “is this directory the kind of structured, authoritative source that LLMs will treat as ground truth?” Different question, different answers, mostly different directories.
Why freshness now beats backlink authority
Old SEO logic: high-authority backlinks compound over time. New AI-discovery logic: recent activity signals trustworthiness more than historic authority. A directory listing updated last month with current pricing, recent reviews, and current photos outperforms a five-year-old listing on a higher-authority site.
This has practical implications. The “set it and forget it” approach to directory listings, which honestly worked fine from 2010 to 2020, actively damages visibility now. Stale listings get demoted in citation patterns. I’ve watched it happen in real time across three clients.
Quick tip: Set a calendar reminder to refresh your top five directory listings every 60 days, even if nothing has changed. Update the “last reviewed” date, swap one photo, tweak one service description. Freshness signals matter more than the size of the change.
The audit cadence every business needs
Here’s what I now build into every retainer:
- Quarterly: Full LLM citation audit (the 47-prompt matrix, or its equivalent for your industry)
- Monthly: Spot-check audit (10 priority queries, 15 minutes)
- Weekly: Review monitoring across the top three citation sources
- Annually: Full directory portfolio review, what to keep, cut, add
The quarterly audit catches drift before it becomes catastrophe. The monthly spot-check catches sudden shifts (LLM training data updates, competitor moves). The weekly review catches reputation problems before they calcify into bad citations. The annual portfolio review catches structural changes: new directories rising, old ones decaying.
Myth: AI discovery is too new and unstable to build a strategy around, so it’s better to wait until things settle. Reality: The test-and-learn approach beats analysis paralysis precisely because the market is moving. Businesses that establish a baseline now will have 18 months of accumulated learning when their cautious competitors finally start. Discovery, as Teresa Torres puts it, works best when scoped to a clear outcome, not delayed until conditions feel certain.
The HVAC client renewed their retainer in February and added a second location to the program in May. The two-van competitor that started this whole investigation? Still cited frequently, still doing well. They didn’t lose; my client just stopped being invisible. That’s how this works now: discovery isn’t zero-sum, but invisibility is, and the businesses that treat discovery as a continuous capability rather than a one-time setup are the ones writing the next decade of customer acquisition. The audit you run next week will tell you which side of that line you’re on.
