Discover how AI platforms perceive your brand. Citable provides AI-powered brand visibility analysis and generative engine optimization (GEO) to improve your digital citability.
Replies
Best
At thousands of prompts per day across models plus persona memory, non-determinism and model drift can make share-of-voice and citation deltas noisy.
Best practice is to run a versioned eval harness: fixed prompt suites, replicated runs, temp 0 where supported, response caching, and calibrated LLM-judge scoring to detect meaningful changes.
How are you normalizing citations (canonical URL dedupe) and isolating persona state from model updates when computing SOV trends?
@ryan_thill great point — we treat SOV/citation trends as an eval problem, not a single run. We run versioned prompt suites with replicated runs, cache outputs, and normalize citations via canonical URL + redirect + tracking-param dedupe so the same page doesn’t inflate counts. Personas are snapshot/versioned (frozen state) so we can separate persona effects vs model updates, and when a model drifts we start a new baseline instead of mixing trends.
@daniele_packard thank you!! 🙌 GEO is different from SEO because the goal shifts from ranking on a list of links → to being the recommended answer inside the model (and ideally cited).
SEO signals are mostly: keywords, backlinks, technical health, CTR. GEO signals are more like: clear entity positioning(“what you are / who you’re for”), source credibility (being referenced by trusted sites), consistent facts + terminology across the web, and content that’s easy for models to quote/cite (comparisons, pricing/benchmarks, specs, use cases).
Report
Interesting take on “AI visibility” as a measurable surface, not just SEO rebranded.
One question though: how do you separate being mentioned from being trusted by agents? As agents start making operational decisions, not just citing sources, the difference between visibility and reliability feels critical.
@gnizdoapp Totally agree — being mentioned ≠ being trusted.
We track those separately: visibility = share of voice / mentions, while trust = “would the model actually recommend you” (measured via preference-style prompts, citation quality, consistency across personas, and whether you’re framed as the “default” choice vs just listed). As agents start making decisions, we think the real metric becomes probability of selection, not just presence — we’re building more of that “reliability layer” into the scoring.
Report
This is awesome and I've been looking for a tool like this for my project and congrats on the great launch! Before committing to paying monthly is there anyway to see (on your platform or off) how often there are prompts related to your product that would appear if it was ranked higher in the AI agent?
Kind of like a pre-qualifier to begin optimizing your outreach through AI agents?
@jake_friedberg Love this question — and yes, that’s exactly the right “pre-qualifier” to ask.
We’re working on an opportunity view that shows the prompt clusters you should be winning (and how often they come up), so you can prioritize the highest-impact queries before spending time optimizing everything. Even today, we can run a quick audit to estimate where demand exists + what prompts you’d likely show up for once you move up.
Report
Congrats on the launch! The shift from pure visibility analytics to visibility + action makes a lot of sense, especially now that AI discovery is more about reasoning than rankings. I like that you’re closing the loop with concrete outputs like content generation and community discovery instead of just dashboards. How you prioritize actions when signals conflict, for example, if different AI engines surface different visibility gaps, how does Citable decide what’s most impactful to tackle first?
@vik_sh Love this question. We prioritize based on impact + confidence: which prompts drive the most demand, where you’re losing to competitors, and how consistent the gap is across repeated runs/personas/models. If engines disagree, we usually start with the “overlap wins” (fixes that improve multiple engines at once), then tackle the highest-value engine for your GTM (ex: Perplexity for research-heavy categories, ChatGPT for mainstream eval, etc.). Goal is always fastest measurable lift, not really the perfect coverage.
Report
very cool! is this agnostic to language? LLM can discover content in different language right? For a korean brand to reach European customers, can you provide insights?
We support 12 countries right now, and content generation works in basically any language. So for a Korean brand targeting Europe, we can test visibility by geo/persona, then help you win locally by finding the exact sources the models trust in that region (press, review sites, communities, directories, etc.) and generate localized content + a distribution plan to get mentioned in the right places.
@abod_rehman Hahaha yes 😂 exactly. ChatGPT gives you the answer, but it doesn’t tell you why that brand showed up...
Report
Hi, Congrats on the launch.
One question I’m curious about: how do you actually evaluate whether these suggested actions are effective? Since generative engines don’t have transparent ranking systems like traditional SEO, what signals or indicators do you use to validate that an action improves visibility, citation likelihood, or model selection?
The core unit we track is persona-based prompt visibility: for a fixed set of real user prompts and personas, how often your brand is mentioned, cited, and positioned relative to alternatives.
Concretely, for any on-page, off-site, or PR action, we measure before/after across:
Mention rate (AI visibility): % of prompts where the brand appears
Citation rate (AI citability): how often the brand is referenced as a source, not just named
We keep prompts, personas, and models fixed, run repeated evaluations, and look for consistent deltas, not single-run wins.
While models don’t expose rankings, they optimize for useful, defensible answers. Actions that improve clarity, authority, and retrievability show up as higher mention and citation frequency across these controlled prompt suites.
In short: Citable validate GEO actions the same way models learn, by measuring how often and how strongly they choose you when real questions are asked.
Report
Yeah! Love the Notion style but even more how you help brands to face their presence in this AI era, which is actually great! Wish you all the best!
This is really cool and makes a ton of sense, considering the direction in which commerce and product research are going. I get how you assess share of voice -- that's clever. But having measured that, how do you help users improve their standing?
@derekattonkotsu Thank you!! 🙏 And yes — measuring SOV is only step 1.
Step 2 is the part we care about most: we turn gaps into specific actions (fix missing citations, improve category comparisons, publish the right “AI-readable” pages, generate content + distribution across channels, and monitor what actually moves the needle). Basically: diagnose → recommend → execute → re-test.
Replies
At thousands of prompts per day across models plus persona memory, non-determinism and model drift can make share-of-voice and citation deltas noisy.
Best practice is to run a versioned eval harness: fixed prompt suites, replicated runs, temp 0 where supported, response caching, and calibrated LLM-judge scoring to detect meaningful changes.
How are you normalizing citations (canonical URL dedupe) and isolating persona state from model updates when computing SOV trends?
Citable
@ryan_thill great point — we treat SOV/citation trends as an eval problem, not a single run. We run versioned prompt suites with replicated runs, cache outputs, and normalize citations via canonical URL + redirect + tracking-param dedupe so the same page doesn’t inflate counts. Personas are snapshot/versioned (frozen state) so we can separate persona effects vs model updates, and when a model drifts we start a new baseline instead of mixing trends.
Cloudthread
Congrats! How different is GEO from SEO? Are there different signals?
Citable
@daniele_packard thank you!! 🙌 GEO is different from SEO because the goal shifts from ranking on a list of links → to being the recommended answer inside the model (and ideally cited).
SEO signals are mostly: keywords, backlinks, technical health, CTR. GEO signals are more like: clear entity positioning(“what you are / who you’re for”), source credibility (being referenced by trusted sites), consistent facts + terminology across the web, and content that’s easy for models to quote/cite (comparisons, pricing/benchmarks, specs, use cases).
Interesting take on “AI visibility” as a measurable surface, not just SEO rebranded.
One question though: how do you separate being mentioned from being trusted by agents?
As agents start making operational decisions, not just citing sources, the difference between visibility and reliability feels critical.
Citable
@gnizdoapp Totally agree — being mentioned ≠ being trusted.
We track those separately: visibility = share of voice / mentions, while trust = “would the model actually recommend you” (measured via preference-style prompts, citation quality, consistency across personas, and whether you’re framed as the “default” choice vs just listed). As agents start making decisions, we think the real metric becomes probability of selection, not just presence — we’re building more of that “reliability layer” into the scoring.
This is awesome and I've been looking for a tool like this for my project and congrats on the great launch! Before committing to paying monthly is there anyway to see (on your platform or off) how often there are prompts related to your product that would appear if it was ranked higher in the AI agent?
Kind of like a pre-qualifier to begin optimizing your outreach through AI agents?
Citable
@jake_friedberg Love this question — and yes, that’s exactly the right “pre-qualifier” to ask.
We’re working on an opportunity view that shows the prompt clusters you should be winning (and how often they come up), so you can prioritize the highest-impact queries before spending time optimizing everything. Even today, we can run a quick audit to estimate where demand exists + what prompts you’d likely show up for once you move up.
Congrats on the launch! The shift from pure visibility analytics to visibility + action makes a lot of sense, especially now that AI discovery is more about reasoning than rankings. I like that you’re closing the loop with concrete outputs like content generation and community discovery instead of just dashboards. How you prioritize actions when signals conflict, for example, if different AI engines surface different visibility gaps, how does Citable decide what’s most impactful to tackle first?
Citable
@vik_sh Love this question. We prioritize based on impact + confidence: which prompts drive the most demand, where you’re losing to competitors, and how consistent the gap is across repeated runs/personas/models. If engines disagree, we usually start with the “overlap wins” (fixes that improve multiple engines at once), then tackle the highest-value engine for your GTM (ex: Perplexity for research-heavy categories, ChatGPT for mainstream eval, etc.). Goal is always fastest measurable lift, not really the perfect coverage.
Citable
@jasonge27 Yep 100% — we can do this.
We support 12 countries right now, and content generation works in basically any language. So for a Korean brand targeting Europe, we can test visibility by geo/persona, then help you win locally by finding the exact sources the models trust in that region (press, review sites, communities, directories, etc.) and generate localized content + a distribution plan to get mentioned in the right places.
Triforce Todos
This is what my brain wants when I Google, but ChatGPT already knows.
Citable
@abod_rehman Hahaha yes 😂 exactly. ChatGPT gives you the answer, but it doesn’t tell you why that brand showed up...
Hi, Congrats on the launch.
One question I’m curious about: how do you actually evaluate whether these suggested actions are effective? Since generative engines don’t have transparent ranking systems like traditional SEO, what signals or indicators do you use to validate that an action improves visibility, citation likelihood, or model selection?
Citable
@blackheart2 Thanks mate! Great question.
The core unit we track is persona-based prompt visibility: for a fixed set of real user prompts and personas, how often your brand is mentioned, cited, and positioned relative to alternatives.
Concretely, for any on-page, off-site, or PR action, we measure before/after across:
Mention rate (AI visibility): % of prompts where the brand appears
Citation rate (AI citability): how often the brand is referenced as a source, not just named
We keep prompts, personas, and models fixed, run repeated evaluations, and look for consistent deltas, not single-run wins.
While models don’t expose rankings, they optimize for useful, defensible answers. Actions that improve clarity, authority, and retrievability show up as higher mention and citation frequency across these controlled prompt suites.
In short: Citable validate GEO actions the same way models learn, by measuring how often and how strongly they choose you when real questions are asked.
Yeah! Love the Notion style but even more how you help brands to face their presence in this AI era, which is actually great! Wish you all the best!
Citable
@german_merlo1 appreciate that a lot!! 💛 We’re obsessed with making AI visibility feel simple + actionable. Thanks for the support!
Tonkotsu
This is really cool and makes a ton of sense, considering the direction in which commerce and product research are going. I get how you assess share of voice -- that's clever. But having measured that, how do you help users improve their standing?
Citable
@derekattonkotsu Thank you!! 🙏 And yes — measuring SOV is only step 1.
Step 2 is the part we care about most: we turn gaps into specific actions (fix missing citations, improve category comparisons, publish the right “AI-readable” pages, generate content + distribution across channels, and monitor what actually moves the needle). Basically: diagnose → recommend → execute → re-test.