Ben Paster @BenPaster

something between a software engineer & marketer Philadelphia, PA Joined July 2009

Tweets

30
Followers

462
Following

566
Likes

594

Ben Paster @BenPaster

a month ago

Many landing pages linked from duplicated top ads has worked well for us too. I think the entity id concept is extended to mean things it doesn't necessarily mean. Meta knows it's the same ad, so: 1) If the same ads end up in the same auction, one of them may be removed from that auction. But there is lots of noise/entropy in the system, so this doesn't _always_ happen. 2) The ad is probably retrieved for the same audience (andromeda). But that's step 1 of the auction, there's several subsequent steps that could filter Ad A but not Ad B, even if they are duplicates.

0 0 0 28 0

View Details

Ben Paster @BenPaster

2 months ago

@karri_tweets @stuartchaney It seems to me that people are building some equivalent of a “what’s my ROAS, by ad” dashboard. Is the market telling SaaS that’s all they actually needed, or is the market just excited they can build for the first time?

0 0 1 99 0

View Details

Ben Paster @BenPaster

2 months ago

PSA: ban your Claude from saying “doing the real work”

0 0 1 66 0

View Details

Ben Paster @BenPaster

2 months ago

What is the incentive to improve optimizers for AI workloads? Anecdotally, it seems that coding workloads are somewhat well optimized (token use is pretty efficient) but most other workloads burn tokens like crazy. Presumably, this would be solvable, unless code is inherently more token efficient than other data? If the optimizers just need a little time to catch up, could that be a bear case on at least some of the AI spending?

0 0 0 47 0

View Details

Ben Paster @BenPaster

2 months ago

why are we jumping to assuming this is bad? what if the worm-scare is responsible for driving a good outcome?

Bryan Cano @BryanECano

2 months ago

This is a DPA (catalog) ad — Chewy did not make the intentional decision to upload this image as an ad. In large catalogs the products selected lean towards those with the highest CTR (not ROAS) so you end up with situations like this. Saw this a lot working on Wayfair and

3 0 19 6K 11

0 0 0 101 0

View Details

Ben Paster @BenPaster

2 months ago

Creative testing really doesn’t work at all. Not even in an A/B test. Divergent delivery is a big issue. arxiv.org/abs/2508.21251

Rok Hladnik @rokhladnik

6 months ago

Creative testing only works if you narrow variables intentionally. Otherwise, you’re just guessing — and reading KPIs after the fact. That’s why so many teams feel “lost” in testing.

6 0 15 100K 1

0 0 0 60 0

View Details

Ben Paster @BenPaster

2 months ago

Question for people who consult with businesses: All these @viktor__com ads are like “Viktor what’s our revenue this week?” Is that a big problem to solve? Do people not know their own revenue (/ ROAS / some other metric that’s directly reported) ?

0 0 1 50 1

View Details

Ben Paster @BenPaster

4 months ago

allbirds would make more money selling their Shopify theme it’s so clean

1 0 2 135 0

View Details

Ben Paster @BenPaster

4 months ago

have you tried making review skill(s) for claude to review its work before it stops? claude.md - run `/review` skill before completing work /review (something like): - load the instructions from {...} - review the landing page and the instructions and create a todo list for each instruction that needs to be reviewed, with context from the original prompt and landing page - start a subagent to confirm each instruction was implemented correctly - if no, update the landing page to comply with the instruction

1 0 0 189 1

View Details

Ben Paster @BenPaster

4 months ago

Re "If it gets a 50% higher CPA than your baseline, how do you know the impressions it drove didn’t cause your offer ad to get a 15% lower CPA than baseline at more spend than before?" do you think it would be actionable to model the relationship between an individual ad's spend vs ad set (or campaign) performance?

1 0 0 181 0

View Details

Ben Paster @BenPaster

5 months ago

Definitely agree, almost certain that will happen. My argument is that a significant percentage of people enjoy shopping. What does the experience look like for them? People who enjoy shopping might also over-index on DTC (tend towards higher disposable income, willingness to explore lesser known brands), though that's just speculation, nothing to support that. i'm sure it depends on product category too, maybe nobody enjoys shopping for paper towels.

1 0 1 28 0

View Details

Ben Paster @BenPaster

5 months ago

I just started trying out @conductor_build. It seems like it did a good job planning/implementing a feature, but also was much slower than native claude code and also used a lot more tokens. is the slower + more tokens thing real or am I imagining things? also it may have written much better code, I didn't review in detail yet.

0 0 1 157 0

View Details

Ben Paster @BenPaster

5 months ago

@binghott I'm sure you can use spend too, but how do you decide what spend level is "good" ?

1 0 0 22 0

View Details

Ben Paster @BenPaster

5 months ago

Proportion of ads that generate a meaningful percentage of total purchases. At least in my experience, it’s a very easy trap to have the team doing busy work creating a lot of ads that don’t matter. (Even if the cpa is good. And spend thresholds are always arbitrary.) There has to be some methodology for tracking if we’re making progress over time. If more ads are selling more stuff, it’s one way (in a toolkit of other metrics) to know if we’re actually making progress or just spinning our wheels.

1 0 0 32 0

View Details

Ben Paster @BenPaster

6 months ago

@KevinSimo123 how do you model the value of each touchpoint for assisted conversions?

0 0 0 27 0

View Details

Ben Paster @BenPaster

6 months ago

@ChereneAubert it’s just so interesting because they’re essentially doing more work, the dashboard was made for them! but I guess ai does the last step of fully synthesizing the information, vs a (good) dashboard the user still had to do the last 10% of the work

1 0 1 19 0

View Details

Ben Paster @BenPaster

6 months ago

@ChereneAubert so am I understanding correctly that you’re not suggesting ai will necessarily provide these end users with better info, but it will allow you to more easily provide them with better info?

1 0 1 21 0

View Details

Ben Paster @BenPaster

6 months ago

@ChereneAubert what have you found people do when they have a question (that can be answered with data)? do they seek out the answer? pre and post slack agent