Barany

building @voxoriahq: marketing data for AI agents. london. sharing what actually works doing marketing with claude code (and what very much doesn't)

[Barany @barany87]

so i'm building @voxoriahq: marketing data for AI agents

right now an agent doing marketing needs a different data vendor for every source, each with its own key, plan and format. exhausting

voxoria is one connection, pay per call. i'll share what i learn here, flops included

[Barany @barany87]

@launch_llama I hope so! 🤞

[Barany @barany87]

how i'd check google AI mode for my category:

1. write the 10 questions a buyer would actually ask
2. have an agent run each through AI mode and normal google
3. list domains that turn up in one but not the other
4. do it again next month and compare

that's it. that's the post

[Barany @barany87]

@gregisenberg On the reviews prompt, i'd also ask it to date each complaint and split the last 90 days from everything older. a top-10 complaint that's mostly from last year is probably fixed already. the ones climbing are the gap worth building into

[Barany @barany87]

@ShamsiSays the MANUAL list is the bit to watch. X replies, reddit comment threads and app store reviews are hard for a model's web search to read, so a lot of the best customer quotes end up on it. worth one run just to see how long that list gets

[Barany @barany87]

@ChrisSimpson The six-city pull is the clever bit. one check i'd add: google localises results to where the searcher is, so "travel agency" searched from inside a city and "travel agency [city]" searched from elsewhere can show different pages. i'd pull both before trusting the word list

[Barany @barany87]

@WalterAIBuilder Love one bot per job. the thing i'd check is what memoh works from: SEO agents get a lot sharper with live google results, keyword volumes and the AI overview for the queries you care about. i'm building @voxoriahq, which gives an agent exactly that, pay per call

[Barany @barany87]

@eshaankansal Step 5 is the one i'd hand to an agent. views ÷ followers per video is easy maths once it can pull an account's recent tiktoks or reels with view counts, then it flags the 10x ones. i'm building @voxoriahq, to automate the pull.

[Barany @barany87]

@sofiabuilds ooh for a couples app i'd search competitor reviews for "my partner", "my boyfriend" or "my wife". they show whether the second person ever actually joined, and my guess is that's where most couples apps quietly stall

[Barany @barany87]

@TheViableEdge 15.4 is an average, so it can hide a lot. a handful of queries sitting at 40+ with loads of impressions drag it down while others are already top 10. i'd export the queries and bucket them 1-10, 11-20 and 21+. my bet is most of the 54 clicks live in the first bucket

[Barany @barany87]

@samfiguresout genuinely curious: did google AI mode pull from the pages that actually rank? when i ran 10 "best software" questions through it, only 22% of its sources shared a domain with the organic results. would love to know if the other five looked the same

[Barany @barany87]

@qadir5000 918 PRs, i can't. on the map grids: save the grid centre, spacing and search term with every run. if any of them change, a drop on the grid can just be a moved grid, and that is not a fun thing to untangle three months later

[Barany @barany87]

@bird_helen67403 really like that you shared the repo. with 3 runs per prompt though, 27.5% and 60% could both be noise. 5 to 10 runs each with the range reported would show which wording gaps are real. logging cited sources would also show if the drift starts in retrieval or the answer

[Barany @barany87]

@hifloecom for the saturation check i'd look at two numbers in the meta ad library: how many advertisers run the angle, and how long the oldest ad has been live. loads of advertisers but nothing older than a few weeks usually means everyone's testing it and quietly dropping it

[Barany @barany87]

@chijuano one filter i'd add: reviews that name another app, like "switched from" or "went back to". they tell you who people are actually comparing you with, which the 1-star rants on their own honestly don't

[Barany @barany87]

@webjuice_ie the second-agent rule is the bit i'd steal. one thing that seems to help: give the checker the brief and the finished work but not the builder's notes. hand it the notes and it tends to follow the same reasoning and just nod along

[Barany @barany87]

@irentdumpsters the second prompt is where it gets a bit hopeful. a GSC export doesn't tell the AI who's ranking above you, so it's guessing. pull the live google results and the maps pack for each keyword first and it's comparing real pages. i'm building @voxoriahq for exactly that bit

[Barany @barany87]

@BowTiedBills checks 3 and 5 need the actual reviews, with dates and owner replies, and claude can't see those from a pasted profile. i'd have an agent pull the google maps reviews first, then run your prompt on real data. (full disclosure, i'm building @voxoriahq, which does that pull)

[Barany @barany87]

@xMezbah honestly this is so me. i'm building @voxoriahq, which lets an agent search X, reddit and google maps reviews in one place. i use it for my own listening every day, and the lesson so far: the search is the easy bit, writing queries that aren't 90% noise is the actual work

[Barany @barany87]

ran 10 "best software" questions through google's AI mode expecting it to cite whatever ranks

it did not

27% of its sources were youtube videos. in normal results for the same searches? 2%

small sample, 1 day, US only. but if your category has no youtube presence, worth a look