Lettertrace measures how often Claude, ChatGPT, and Gemini mention your company. But there's a catch: it's free, developer-first, bring-your-own-key, and open source.
Replies
Best
The question I have not seen asked yet is what makes the number stable enough to trend. Ask the same model the same thing twice and you get different answers, and a silent model update moves your baseline without telling you. So a line going up is mixing your own work with drift you cannot see.
Cheap fix, and I think your architecture already allows it. Measure a control set in the same run, a few competitors or unrelated brands you are not touching. Then drift shows up as a common shift across all of them and you can subtract it out. Being BYOK and open source also means you can pin model versions and re-run history, which the 250 a month tools cannot really offer.
Report
If a company shows up in 15/24 answers, do you know whether that correlates with anything downstream - inbound traffic, signups, or even people mentioning they found you through AI? That second number seems harder to get than the visibility score itself. Congrats on the launch!
Report
Awesome product! I use it today :) Congrats on the launch Mathew!
Report
Wow, interesting! Will it show which sources the AI was trained on, so I can add information about us to those sources? That’s one of the key things for GEO optimization...
Report
"AI search is effectively measured by a crafty cron job that orchestrates a bunch of model provider calls" is the sentence that should be at the top of the page. Naming the thing you are commoditising is the whole pitch.
One property of BYOK worth putting in the docs before a user finds it themselves: everyone queries from their own account, which increasingly means their own model version, their own region, and whatever personalisation has accumulated there. Two people running Lettertrace against the same company on the same afternoon can get materially different share of voice, and neither of them is wrong. So the number is comparable to itself over time only if the model underneath held still, and it never does.
Not a flaw, it is the shape of the problem. But "mentioned in 15 of 40 prompts" reads like a measurement and behaves like a sample, and that gap is where this category loses trust.
Cheap thing that would help more than it sounds: pin the model version into the output line.
Replies
The question I have not seen asked yet is what makes the number stable enough to trend. Ask the same model the same thing twice and you get different answers, and a silent model update moves your baseline without telling you. So a line going up is mixing your own work with drift you cannot see.
Cheap fix, and I think your architecture already allows it. Measure a control set in the same run, a few competitors or unrelated brands you are not touching. Then drift shows up as a common shift across all of them and you can subtract it out. Being BYOK and open source also means you can pin model versions and re-run history, which the 250 a month tools cannot really offer.
If a company shows up in 15/24 answers, do you know whether that correlates with anything downstream - inbound traffic, signups, or even people mentioning they found you through AI? That second number seems harder to get than the visibility score itself. Congrats on the launch!
Awesome product! I use it today :) Congrats on the launch Mathew!
Wow, interesting! Will it show which sources the AI was trained on, so I can add information about us to those sources? That’s one of the key things for GEO optimization...
"AI search is effectively measured by a crafty cron job that orchestrates a bunch of model provider calls" is the sentence that should be at the top of the page. Naming the thing you are commoditising is the whole pitch.
One property of BYOK worth putting in the docs before a user finds it themselves: everyone queries from their own account, which increasingly means their own model version, their own region, and whatever personalisation has accumulated there. Two people running Lettertrace against the same company on the same afternoon can get materially different share of voice, and neither of them is wrong. So the number is comparable to itself over time only if the model underneath held still, and it never does.
Not a flaw, it is the shape of the problem. But "mentioned in 15 of 40 prompts" reads like a measurement and behaves like a sample, and that gap is where this category loses trust.
Cheap thing that would help more than it sounds: pin the model version into the output line.