Arnav Sharma retweeted
most men wait to feel like a winner before they act like one
that's the wrong order
the brain doesn't hand out confidence in advance. it builds it backwards, through action, through repetition, through showing up when nothing feels certain
the man who trains alone at 6am when no one is watching
the man who keeps his word to himself when no one would know if he didn't
the man who stays calm when everything around him is falling apart
he's not calm because his life is easy
he's calm because he's been practicing being that man for years
winners don't think differently because they won
they won because they started thinking differently before anything in their life reflected it
your circumstances don't have to change first
your mind does
act like the man you're trying to become
long before the results show up
because by the time the results arrive
you'll already be him
Price per token is the wrong number to compare (via @timjcarden).
A cheaper model that needs 3 tries costs more than a pricier one that gets it right the first time.
Same with an AI receptionist. The number that matters isn't cost per minute. It's cost per booked job.
And most of this year's gains aren't coming from smarter models. They're coming from the boring parts: checks, retries, logs, and an alert when something fails.
Which boring part would you fix first in your setup?
3
A pest control company's website form said "submission failed." Their Yelp leads never reached the CRM. Nobody noticed for about a week (via @aireceiptswork).
That's the leak nobody buys an AI receptionist to fix.
The test I'd run once a month. It takes 10 minutes.
Send 1 fake lead through every source. Website form, ad form, Yelp, a missed call.
Check each one lands in the CRM, tagged by where it came from.
Check someone actually got alerted.
Then set one alert for any new lead nobody has touched within an hour.
Most businesses have never done the first step.
How many lead sources does yours have?
13
276 calls a month is the easy part to measure. the number i'd want next to it is how many got answered, and what happened to the ones that hit voicemail. that's usually where the profile's work leaks out
We set up a Google Business Profile for a plumbing company and within 7 months it was doing 276 calls a month.
Owners obsess over the wrong parts of their profile.
Here’s every section, graded by what it actually does.
THE RANKING LEVERS
13
honestly a 6 right now, heading to an 8. the first move isn't even ai: a text back within a minute when a call gets missed. most small businesses lose more jobs to voicemail than to a rival's ai
How much trouble is the average small business in if their rivals go all-in on AI and they don't?
1 to 10. Be brutally honest.
What is the first thing you'd do to ensure they can compete?
6
where i'd use this in a client build: not on the live call, the latency kills it. on the step after, where the agent writes to the crm or books the job. one wrong booking costs more than running it 3 times
The easiest way to make an agent 30 to 50% more accurate is to run it 3 times
Give the same task to a few agents in parallel and let a judge pick the best answer. It was the biggest accuracy jump of the 5 agent patterns in my blog
The catch is it costs 2 to 5x. You pay for every agent and then the judge on top. Speed barely changes because they all run side by side, so you just wait for the slowest one plus the judge
The gains fall off fast though. 5 agents might only be 5 to 10% better than 3 and cost about 67% more. So 2 or 3 is usually plenty
Feels like it only makes sense where a wrong answer is expensive. For simple stuff the extra cost isn't worth it
Read more at manthanguptaa.in/posts/agent…
8
Arnav Sharma retweeted
This is why the human checks remain vital to success, even when using AI.
AI isn't at the point where it can reliably handle everything *without supervision.*
You have to be the overseeing manager on duty, lol
"The agent marks its own homework." Best line I read today, from @mardehaym's post on AI coding factories.
It fits client automations just as well. The workflow says "sent". The AI agent says "booked". Nobody checks whether the text arrived or the slot exists.
His math: if a tool saves 20 hours a week but takes 40 hours of your best people to check, you're behind.
Who checks your AI's work, a person or another AI?
13
"Audit what ChatGPT says about local businesses, then charge monthly to fix it" is the new agency pitch.
The question in the replies was fair: what is the monthly fix?
For a local service business, I'd say it's 4 boring things.
The Google Business Profile matches the website exactly. Name, address, hours, services.
A page for each service and each town, written the way customers actually ask.
Schema on the site, so the AI gets a cheat sheet instead of guessing.
A steady flow of new reviews, which is the part a CRM can run on its own.
None of it is new. It's local SEO with a fresh reason to finally do it.
What would you add?
16
HighLevel's Managed Agents can now run on a schedule.
You set it up by just asking: "Every weekday at 8 AM, summarize my new leads and inactive deals." The agent runs it and posts the result back in the same chat.
The details:
Up to 10 active routines per person, per agent.
The shortest interval is 1 hour.
It can use Computer to browse websites as part of the task.
Every run is billed like any other Managed Agents run.
That last line is the one I'd check first. A daily summary across 20 client sub-accounts is 20 runs a day.
This is close to the "owner gets one text every morning" setup I've been planning to build. What would you have it summarize first?
14
for a local service business the schema that matters is the boring one: the exact name, address, hours and service area, matching the google profile word for word. a mismatch there is exactly the guess you were trying to stop
Schema is code that makes your business machine readable.
Think of it as giving AI a cheat sheet instead of making it guess.
15
82 is a lot to keep honest. the ones that earn their keep in an agency setup are boring: the missed-call text-back, the follow-up nobody remembers to send, and the review request after a job. if you could only keep 3, which ones?
9
Marc Lou pulled numbers from 6,829 startups on TrustMRR.
Made any money in the last 30 days: B2C 63.6%, B2B 54.7%.
Made $10k or more: B2B 4.2%, B2C 2.4%.
So B2C gets a first dollar more often, and B2B is almost twice as likely to reach $10k.
My own small SaaS is B2B, forecasting for cleaning franchises, with 3 to 4 clients. Slow to win each one.
Which would you build first if you were starting tonight?
13
HighLevel says your sending domain is verified. That's where most people stop.
3 things that still decide whether your emails land:
DMARC. It doesn't connect on its own. You add it by hand as a text record in your DNS, then verify again.
Warm-up. HighLevel added Email Domain Warm-Up to Labs this week. A brand new domain blasting a list on day one is how you end up in spam.
The list. Even a perfect setup goes to spam if the people on it ignore, delete or report your mail. Inbox providers watch that.
I haven't sent a big campaign from HighLevel yet, so this is the checklist I'm running before I do.
What fixed your deliverability, the setup or the list?
8
118 of 168 hours is the stat i'd lead with too. the cheapest version isn't even a voice agent. a missed-call text within a minute catches the sunday-night leak before the homeowner tries the next number. mine landed in 41 seconds
A shop open 8 to 6 on weekdays is closed 118 of the week's 168 hours.
Water leaks don't know that. It happens Sunday night and the homeowner calls whoever picks up.
If I ran a small plumbing company, the after-hours phone is the first thing I'd hand to AI.
10
most highlevel agencies serving local businesses sit in bucket 1 and don't know it. setting up a text-back and a review request is execution. the ones that last will own the outcome, booked jobs per month, not the setup
The demand for retention marketing has never been higher but I still see doomers talking about AI replacing all agencies. Here’s my take.
Retention agencies live in 1 of 3 buckets:
Bucket 1: $2,000 to $5,000 per month retainers. Clients with this budget are making decisions primarily on bandwidth/execution and cost. These will predominantly be replaced by AI tools. They were never strong with strategy. They execute inside strict systems and processes built to protect their margins.
Bucket 2: $5,000 to $10,000 per month retainers. Better quality services but the margins still aren't strong enough for high-level strategy work (under the current model) and any decent amount of volume. Still mostly junior teams, offshore support, and rigid processes.
Bucket 3: $10,000 to $15,000 per month and above. This is where brands are paying for deep insights, market perspective, and retention specialists who operate as an extension of their team. Email and SMS, but also subscription optimization, loyalty programs, physical mailers, mobile apps, etc. Experienced teams, extra bandwidth. Top tier creative. Seniority and strategy to match.
The agencies in bucket 3, if they continue utilizing AI as leverage, will pull away. They're building proprietary tools, data-backed strategies, and systems for the bigger retention problems. They will likely be able to offer higher value and potentially lower their prices and squeeze the offers in bucket 2.
Bucket 1 is hired to execute and will be replaced.
Bucket 2 is hired to solve problems and will get squeezed eventually.
Bucket 3 is hired to solve for outcomes and will pull away.
8
A voice agent can pass every test and still be broken on real calls (via @manthanguptaa).
The transcript reads fine. The right tool got called. Then you listen to the recording: it talks over the caller, mangles their name, and leaves 3 seconds of dead air before confirming.
"Friday at four" heard as "Friday at floor" looks like the model got confused. The model was just handed the wrong words.
His rule: someone listens to 10 to 20 real calls every week.
Do you test your voice agent by reading transcripts, or by listening?
1
8
Google's Gemini can now phone local businesses for its users. Free, starting in the US.
In Google's own launch example, it called 2 car dealerships at the same time. Only one of them got the appointment. Booked in under five minutes.
Now picture that for a plumber or a cleaner.
The customer never hears your voicemail. Their AI just moves on to the next number on the list.
If an AI called your business right now, would anyone pick up?
1
13
holding the stage move until the data is in is the right call. curious whether you did it with a workflow condition or custom code. my guess is most workaround stages exist because the required fields were never enforced in the first place
Does your CRM have stages your team only uses as workarounds?
I'm rebuilding GoHighLevel for an insurance advisory firm. First piece is live: every stage move is held until the right data is saved and checked 👇
#GoHighLevel #GHL #CRM
7
if the ai answer can't be tracked, the call can. add "how did you hear about us" to the call flow or the missed-call text, and tag the answer in the crm. a month of that gives you a number for ai-sourced calls that search console never will
I used to be an AI search doubter.
AI was clearly where search was headed, but I didn't think it was sending home service companies real calls yet.
Now I think it's bringing in a good share of your calls... but almost nobody can track them.
Search "best roofer in [your city]" or "best plumber near me" and look at the top of the page. For a lot of these searches, Google's AI Overview now sits above the map pack, and it pushes the organic results even further down.
A roofing owner we work with told me two homeowners called him this summer straight from an AI answer. Someone on my team pointed out what a lot of homeowners do instead: grab a few names from the AI answer, then look those companies up on Maps and read the reviews.
Tracking it is the hard part:
> There's no tracking number for an AI Overview, and no UTM tag on the click.
> Search Console shows how many times your site appears in AI Overviews and AI Mode, but not the clicks. Those stay mixed in with your regular search numbers.
> ChatGPT's new Business Hub shows clicks from ChatGPT to your site, which is more than Google gives you for its own AI answers.
Whether you call it AEO, GEO or what I call it, search everywhere optimization, it deserves real budget now.
The good news is that a lot of what ranks you in the map pack helps you show up in AI Overviews too, and the multi-location playbook is as strong as ever.
If someone in Plano or Allen searches for the best roofer and your only Google profile has a Dallas address, you're going to have a hard time showing up. Real offices with their own profiles across the metro give you a shot in each of those towns, on the map and in the AI answer above it.
7















