Doug Manoel

Building small AI tools on the side. right now: measuring how often chatgpt, gemini and perplexity name local businesses. free check below.

[Doug Manoel @DougManoel89]

made a free check for local businesses

type your business and city. it asks perplexity 5 questions a customer might ask, 3 times each, and counts how often you get named out of 15

takes ~40s. one engine only, so it's a first look not a verdict

https://t.co/Lr0xRsSgai

[Doug Manoel @DougManoel89]

Same reason my local checker asks every question 3 times. One run tells you almost nothing.

https://t.co/38EkrUPkPA

[Doug Manoel @DougManoel89]

Best part of Anthropic's subagent guide is the warning.

Folding one subagent into the main loop broke a process that had been passing. It only showed up in 1 of 2 runs.

So the lesson isn't "delete your subagents." It's run each eval more than once. https://t.co/C5Ar46Niy0

[Tak 🦞 @cherry_mx_reds]

Anthropic is now recommending that companies REMOVE unnecessary subagents from their AI systems.

This is a pretty significant shift in agent architecture, with their team finding that a single agent equipped with skills can frequently outperform complex multi-agent systems, and one customer seeing roughly 2x faster execution.

The emerging rules of thumb:

→ Skills: Teach the agent HOW to do something, including procedures, expertise, and instructions.

→ Tools / MCP: Give the agent the ability to DO things through APIs, actions, and live data.

→ Subagents: Use for parallel work, isolated context, or tasks a single agent struggles with.

The interesting part is progressive disclosure. Instead of spawning specialized agents or stuffing everything into a giant prompt, one agent discovers and loads skills and tools as needed.

Their advice is to fold existing subagents into the main loop, turn their instructions into skills, expose their actions as tools, and run evals to see if performance holds up.

As models have gotten smarter, improving agent architecture is becoming less about what you can add and more about what you can delete.

https://t.co/DNq60Skkmc

[Doug Manoel @DougManoel89]

@DanHaslam @yext 4% agreement across four engines is wild. so 'are we visible in AI' is really four different answers. thanks for the link, reading it now

[Doug Manoel @DougManoel89]

@DanHaslam @yext yeah that's where i landed too. blended for the client report, per-engine for the actual work

in the one market i ran, the leader was in 31% of gemini answers and 0% on perplexity. a single score would've hidden that

[Doug Manoel @DougManoel89]

@martincoderra nice, per engine is what i'd want too. do you also track which sources each engine cites, or just mentions?

[Doug Manoel @DougManoel89]

@alexgroberman if you want to see your own number, i made a free check. perplexity only, 15 answers, ~40s

https://t.co/fkFyrW2l3U

[Doug Manoel @DougManoel89]

@alexgroberman matches what i saw in one US dentist market i ran

brightlocal had Maps naming the tracked business in 66% of searches. AI surfaces only 32-38%

in my sample the market leader showed up in 31% of Gemini answers. Perplexity? 0%

you have to check each engine on its own

[Doug Manoel @DougManoel89]

@AlDev0 @AmanShips true. for "best plumber near me" there's nothing to summarize, it has to name someone

so the real question is who it names and why. that's the part i'm trying to measure

[Doug Manoel @DougManoel89]

@DanHaslam @yext yeah one run is just a snapshot. the trend is the useful part

still curious tho, are partners asking for per-engine numbers or is it one blended "AI visibility" score for now?

[Doug Manoel @DougManoel89]

putting Skills in the same foundation as MCP makes sense

70+ platforms already read the format. and google, microsoft and openai engineers sit on the steering committee

so if you write SKILL.md files they're not stuck with one vendor anymore. anyone porting theirs yet? https://t.co/wCL3okgGem

[ClaudeDevs @ClaudeDevs]

We're donating Agent Skills to the Agentic AI Foundation (@AgenticAIFdn), the same foundation we donated MCP to.

Since December 2025, 70+ agent platforms have integrated the standard, including ChatGPT, Cursor, GitHub Copilot, and VS Code.

Here's what that means for you:

[Doug Manoel @DougManoel89]

@neuralloot thanks, gonna read it tonight

so for gemini a city in the prompt isn't enough, it'd need different network exits. honestly the api vs app gap worries me more

[Doug Manoel @DougManoel89]

@AlDev0 @AmanShips makes sense. informational is basically AI overview territory now

are you seeing transactional hold up in chatgpt answers too, or mostly on google?

[Doug Manoel @DougManoel89]

@neuralloot wait, so Gemini follows the network and not the declared city? that changes how i'd set up the multi-city test

which paper is this btw

[Doug Manoel @DougManoel89]

@AlDev0 @AmanShips grouping by intent is smart, hadn't thought of that.
gonna try it on the next run. which intents do you usually split into?

[Doug Manoel @DougManoel89]

@neuralloot Fair. My 8% is one location, one day, 5 runs per question. It works as a baseline for that city, not as a general number. Rerunning the same questions from more locations is the next thing I want to test.

[Doug Manoel @DougManoel89]

@AlDev0 @AmanShips Yeah, and it's easy to miss because the overall number looks fine. I only caught it by breaking mentions down per question. Do you report AI citations to clients per question or as one score?

[Doug Manoel @DougManoel89]

@SamuraiPhonk Same. And the ask only works if it's specific: who it's sending to, what it's paying, what gets deleted. A vague "ok?" gets approved on autopilot.

[Doug Manoel @DougManoel89]

I use Grok Bot every day, so read this with that bias.

Claude Duet and Grok Bot share the core idea: an agent on its own cloud computer that asks before acting.

The difference on paper: Duet is one assistant. Grok Bot is a team of Bots that message each other.

[Doug Manoel @DougManoel89]

@nbgreene14 Agreed. Maps gives you a position; AI gives you a probability. The closest thing to "where you rank" is a share: same questions, asked several times per engine, count how often you're named. Noisy per answer, steadier as a rate.

[Doug Manoel @DougManoel89]

@DanHaslam @yext Thanks. The caveat I'd put on that slide: one market, one day, 25 questions. The split is real in that sample, but it's a baseline to rerun with the same questions, not a verdict on any engine. Are Yext partners asking for per-engine numbers yet?