Stop tracking AI traffic. Track these 3 things instead

conversionninja

Ranking in ChatGPT is the shiny new obsession, but AI referral traffic is a hollow vanity metric. The citation does its job whether or not anyone clicks, because the buyer absorbs the recommendation right there in the answer.

So here's what actually goes into my reports:

Share of voice. Not "were we mentioned" but how often compared to competitors, on the same fixed set of buyer-intent prompts, month over month.
Prompt coverage, and grade every answer instead of just counting it. My scale runs: absent, mentioned, cited, recommended, described incorrectly. That last one should terrify you, because an AI confidently describing your product wrong does more damage than silence.
Framing and sentiment. Is the model calling you "best for X" or burying you at nine in a list of ten?

One practical note: if your prompt set shifts between scans, the trend line is garbage. Freeze it and version it. Learned that the hard way after reading a case study from HumansWith.AI showing how changing prompts mid-experiment turned the whole dataset useless.

Has anyone here moved these numbers on purpose and isolated which change actually did it?

retention-expert

Agreed. Same thing happens in retention work. People obsess over flashy dashboard metrics while the signals that actually predict churn sit ignored. Knowing what customers do after they buy beats counting AI referral hits any day.

adcraft

"Described incorrectly" is the worst outcome on the board, honestly. Being absent means the user might find you somewhere else. Being wrong means they walk away confident in garbage about your brand. Nobody budgets for that problem, and it's the one actually costing revenue.

Would push one step further too: stop reporting AI visibility as one blended number. Same exact prompt, brand crushing it on Perplexity, nowhere to be found on ChatGPT. Seen it repeatedly. One aggregate figure hides exactly which surface is broken. Track per model or you're optimizing against a fake average.

socialselling

That "described incorrectly" bucket keeps me up at night. You can optimize visibility all day, but when an AI confidently hallucinates facts about your product? that's actively working against you - and most brands have zero clue it's happening because nobody's checking.

we caught a few outright fabrications by running regular prompt audits across our key categories. Fixing the source material isn't instant either - different models pick up corrections at wildly different speeds. Some update in days, others take months

harper-local

honestly? most of that "AI traffic" noise doesn't move the needle for local clients. We track what pays the bills - calls from the Map Pack, direction requests, and whether those reviewers actually show up. Built a simple dashboard pulling GBP insights, CallRail, and a citation audit sheet. Updates weekly. takes ten minutes to review.

What's the actual conversion metric you're optimizing for?

fieldevent

this hits close to home. we run the same event activation playbook across multiple shows each quarter, and the day-to-day variance in booth traffic, demo requests, and qualified conversations often dwarfs the quarter-over-quarter trends we're trying to measure.

Learned the hard way that if you don't baseline the noise first - run the exact same setup two days in a row at the same venue - you end up optimizing for randomness. weather, competing keynotes, badge pickup delays... the "uncontrollables" move the needle more than most program changes.

Now we treat the first run as a calibration round. Measure the spread, understand the floor, then you can actually tell when a new demo station or speaking slot genuinely moves revenue. Otherwise you're just writing case studies about luck

chatbox

That trio is exactly what we ship on retainer engagements - frozen prompt set for share-of-voice, the absent→mentioned→cited→recommended ladder for answer quality, and list-position framing for visibility context. Changing prompts mid-stream absolutely kills the trend line, seen it happen too many times.

On the "moved it on purpose" question: yep, by cleaning up the source layer. We pull the URL-to-answer mapping through seoforgpt, flag the pages feeding weak or hallucinated responses, then push targeted KB updates or new supporting assets. Authoritative sources usually show answer shifts inside two to six weeks.

Process looks like:

Lock a representative prompt set (20-50 queries) and never touch it again
Run weekly extractions via seoforgpt - captures citing URLs, snippet text, list order
Score each answer against the four-state ladder, roll up to domain-level SoV
Cross-reference citing URLs in GSC and Ahrefs to verify index status and authority signals
Prioritize content fixes where high-value queries cite thin, outdated, or competitor pages
Re-measure at the two-week mark, if no movement, audit crawl budget and internal linking to the updated assets

The frozen prompt set is the control. Everything else is just noise without it.

churnkiller

ran into the exact same prompt drift problem. what saved my sanity was holding a control set of unrelated queries next to the frozen set, because model updates alone will shift your numbers. if the control queries drift too, it's the model, not you