Agencies
How Do I Know If My AEO Work Is Doing Anything?
Jillian Oco, CMO— AI-assisted and reviewed prior to publication.
Track citation frequency monthly across the four assistants your clients ask, measure changes over 30 days minimum to separate signal from algorithmic noise, and remember that one check is a snapshot. Real measurement requires a baseline, time, and an understanding that retrieval-based assistants like Perplexity update instantly while training-based ones like Claude can take months to reflect your work.
What to measure monthly
Monthly tracking should focus on citation frequency, share of voice, and how often your brand appears in AI-generated answers (airops.com) when someone asks who to hire in your clients' categories. Track AI citations and mentions, share of answer across key queries, brand presence, and assisted conversions from AI-influenced sessions (tryprofound.com).
The simplest starting point is measuring whether ChatGPT, Gemini, Perplexity, and Claude name the business when asked, and which competitors are mentioned alongside them (visible.seranking.com). The free AI-visibility check (/check) shows this for a single point in time. Monthly measurement means running the same queries, in the same locations, on the same day each month, and recording both presence and absence.
Visibility is measured across four dimensions: mentions, citations, sentiment, and position in AI-generated responses, with metrics refreshed weekly (digitalapplied.com). For agencies selling AEO, that means at minimum one check per month per client, saved in a spreadsheet, with the exact query, the assistant, the date, and whether the client was named.
What does noise look like?
Statistical significance helps determine whether a change in metrics is a real effect or just random variation, distinguishing real effects from normal week-to-week wobble. Search demand fluctuates, competitors publish content, and Google reshuffles results. The same volatility affects AI answers.
A 2% conversion rate increase needs 15,000 sessions, roughly three months for mid-traffic sites, before declaring victory (improvado.io). AEO is quieter. A single citation gain might come from a lucky refresh. A single loss might come from an assistant's updated training data that has nothing to do with your optimization.
Statistical significance tells you that a gap is a real effect and not random week-to-week variance (seocrawl.ai). Without it, you risk crediting your work for a change that was already coming or blaming yourself for one outside your control. The practical minimum is 30 days of comparison, with most tests requiring at least 2 to 4 weeks of data collection, though complex experiments may need longer.
Why is a single check just an anecdote?
One snapshot tells you where you stand today, not whether the work is doing anything. A result without a control group or time series is directional evidence, not statistical proof. That matters when you are reporting to a client who expects measurable return on their spend.
The difference between a check and a measurement program is repetition. You need a baseline taken before you optimize, monthly checks afterward, and enough time for the assistants to re-index or retrain. Do not measure AEO with screenshots alone. Track mentions, citations, traffic, sentiment, and conversion quality.
The guide (/guide) explains the mechanics of optimization, but measurement determines whether that work moves the number. A single check is useful for diagnosis, a series of checks is the only way to measure impact.
How retrieval differs from training
Perplexity and Gemini search the web natively for real-time information, while training-based models like Claude Opus 5 rely on data through May 2026 and GPT-5.6 through February 2026. The difference in how they source answers changes how quickly your optimization shows up.
Retrieval-augmented generation (RAG) retrieves relevant documents at query time and feeds them into the model as context, allowing AI to generate more accurate, current, and domain-specific responses. Perplexity is built search-first, while Grok operates in real-time when web or X search tools are enabled. Optimize for Perplexity and you can test the result the same day.
Training-based assistants freeze knowledge at a cutoff. The training data cutoff is the outer edge of material the model was trained on, while the reliable knowledge cutoff is the date through which knowledge is actually dense and dependable, with gaps reaching five months on some models. That means citation changes lag your optimization by weeks or months, depending on when the assistant next retrains.
For agencies, this means setting client expectations around timeframe. Work targeting Perplexity or Gemini can show movement within 30 days. Work targeting ChatGPT or Claude without real-time search may take a full training cycle, and you will not know the schedule in advance.
ActiScore tracks all four on ActiScore's pricing (/pricing) tiers starting at the free plan, running the same queries your prospects ask and showing which competitors the assistants name. The difference between a gut check and a measurement plan is recording the result, repeating it monthly, and comparing the trend line rather than any single point.
Common questions
What should I measure monthly for AEO?
Monthly tracking should focus on citation frequency, share of voice, and how often your brand appears when someone asks who to hire in your category. The simplest starting point is measuring whether ChatGPT, Gemini, Perplexity, and Claude name the business when asked, and which competitors are mentioned alongside them. Run the same queries, in the same locations, on the same day each month, and record both presence and absence in a spreadsheet with the exact query, the assistant, the date, and whether the client was named.
How long does it take to see results from AEO work?
The timeframe depends on which assistant you are optimizing for. Retrieval-based assistants like Perplexity and Gemini can show movement within 30 days because they search the web in real time. Training-based assistants like ChatGPT and Claude without real-time search may take a full training cycle of weeks or months before citation changes appear, since their knowledge freezes at a cutoff date. The practical minimum for any measurement is 30 days of comparison to separate real signal from algorithmic noise and normal week-to-week variance.
Why is one visibility check not enough to measure AEO?
One snapshot tells you where you stand today, not whether the work is doing anything. A result without a control group or time series is directional evidence, not statistical proof, and that matters when reporting measurable return to a client. The difference between a check and a measurement program is repetition: you need a baseline taken before you optimize, monthly checks afterward, and enough time for the assistants to re-index or retrain. A single check is useful for diagnosis, but a series of checks is the only way to measure impact.
What does noise look like in AEO measurement?
Search demand fluctuates, competitors publish content, and algorithms reshuffle results, creating normal week-to-week volatility that affects AI answers. A single citation gain might come from a lucky refresh, and a single loss might come from an assistant's updated training data that has nothing to do with your optimization. Statistical significance helps determine whether a change is a real effect or just random variation. Without tracking over time, you risk crediting your work for a change that was already coming or blaming yourself for one outside your control.
How do retrieval-based and training-based assistants differ for AEO?
Retrieval-based assistants like Perplexity and Gemini search the web natively for real-time information, so your optimization can show up the same day you publish. Training-based models like Claude Opus 5 and GPT-5.6 rely on data frozen at a cutoff date, meaning citation changes lag your optimization by weeks or months depending on when the assistant next retrains. For agencies, this means setting different client expectations around timeframe: work targeting Perplexity or Gemini can show movement within 30 days, while work targeting ChatGPT or Claude may take a full training cycle.