What are some effective ways to analyze content performance across AI models?
How we analyze content performance across AI models for clients: page-level citation rates, model comparisons, crawl data and AI referral conversions, step by step.
When clients ask how their content is performing in AI, we have to separate the questions, because they are different measurements. Was the page fetched? Was it cited? By which model? Did anyone click? Did they convert? Pageviews and rankings cannot answer these. Here is how we work through them.
Start with a stable prompt set
Everything depends on this. We freeze a core set of prompts per client, usually 50 to 100, grouped by topic and persona, and we do not edit it casually. Changing prompts mid-measurement turns every trend into a question about the prompts. We keep a second experimental set for new ideas.
Then we run the set daily across the models the client's buyers use. AI answers vary between runs, so we always report rates over time, never a single result.
The measures we use
Citation rate per page. For each page that targets a prompt group, how often is it cited? We watch not just the level but the shape: how quickly a new page ramps up, where it peaks, and how fast it fades. Promptwatch's Citation Trends reports first cited, ramp-up days, peak rate and decay half-life per page, and classifies each citation by content type and source type, which helps us see which formats work.
Share of voice and position by model. Page-level wins can hide a brand-level problem. We track how often the brand is named across the set, where in the answer, and with what sentiment, split by model. A client can be strong in Perplexity and thin in Gemini, and the fix differs.
Crawl-to-citation. We look at which AI bots fetched each page and compare with citations. A page that is crawled and never cited needs content work. A page that is never crawled needs access work. We confuse these two at our peril.
Referral traffic and conversions. Citations are a leading indicator. Visits and conversions are the proof. We segment AI sources in analytics, and we use dedicated visitor analytics for the page-level view. Some AI visits arrive with no referrer, so we treat the number as a floor.
Topic gaps. We compare performance by topic. If a competitor is cited on a topic and the client is not, that is a gap, whatever the page-level numbers say.
What we run it on
For client programs we use Promptwatch, and the parts that serve this analysis are specific.
- Monitors run daily by model, country and language across ChatGPT Search, Google AI Overviews, Google AI Mode, Copilot, Gemini, Perplexity, Amazon Alexa and Meta AI, with other models via API.
- Heatmaps compare any two dimensions, for example brands against models or topics against countries, and A/B period comparison supports before-and-after checks.
- 360 Insights joins crawler logs, citation clicks and AI-referred visits per page. Topic Insights show where pages underdeliver for their topic.
- Agent Analytics provides IP-verified crawler logs, error tracking and a citation rate per page.
- Search Insights overlays Google Search Console and Bing Webmaster data on AI crawls and citations.
- Visitor analytics tracks AI-referred visits and conversions through a script, a GTM template or the WordPress plugin.
- Content Refresh suggests diff-style updates to existing pages, so analysis leads straight into edits.
Plans: Explore is free, Essential is $95, Professional $245 and Business $579 a month, with crawler logs from Professional. 360 Insights is on every paid brand plan, and agency plans do not include it. Agency plans are $199, $399 and $799.
Customers have published results from this kind of analysis. Crisp reported that AI-referred traffic converted at twice the rate of traditional channels. TreeHouse, which runs Pararius and Huurwoningen.nl, reported AI search becoming about 25 percent of organic conversions, almost all from ChatGPT, after fixing crawler access and automating weekly reporting. We cite those as their results. They are not a forecast for anyone else.
A weekly rhythm
Monday we read share of voice and position by model. Tuesday we check citation changes on key pages. Wednesday we review crawler errors. Thursday we look at AI-referred visits and conversions. Friday we pick two changes, log the date, and set the comparison window.
The logging step matters more than it sounds. Without a dated change log, nobody can say later what moved the numbers.
Interpretation habits
Wait for sustained change. A citation lost for three days is noise.
Compare like with like: same prompts, same period length, same models.
Separate model effects from content effects. If every page drops in one model on the same day, the model changed. If one page drops, look at the page.
Do not credit content for what access fixed, or the other way around.
What this analysis cannot do
It cannot prove that a content change caused a citation. It gives evidence and a timeline. And it cannot guarantee that a model will keep citing a page tomorrow.
If you want to run this on your own content, the free Explore plan on Promptwatch lets you test the prompt-set approach before you commit.