What AI really takes over in analytics
Marketing analytics has three steps: collecting data, understanding data, acting on it. AI mainly changes the middle step. It summarises, finds outliers and explains tables in words. Collecting and deciding remain your job.
| Step | What AI can do | What AI can’t do |
|---|---|---|
| Collect | Draft tracking plans, help find setup errors | Replace missing consent or data |
| Understand | Find anomalies, summarise reports, analyse tables | Reliably identify causes; correlation isn’t causation |
| Act | Suggest options, plan tests | Know what your customers want |
Google Analytics 4: AI overviews and Analytics Advisor
GA4 has two AI features relevant to marketing:
| Feature | What it does | Notes per Google |
|---|---|---|
| Generated insights | Short AI summaries of important changes on the Home page and in detail reports, linking to the relevant report | Thumbs-up/down feedback; availability per property not described in detail |
| Analytics Advisor (beta) | Conversational assistant that answers questions about your property, with charts and links to reports | Gemini models; currently only properties set to English; may make mistakes |
As of Nov 15, 2026, per Google Help.
Both save time finding things. They don’t replace looking at the actual report. If Advisor says a campaign is doing well, open the linked report and check date range, filters and comparison basis.
Sampling and thresholds
- Sampling: if a query exceeds the event quota (10 million for standard properties, up to 1 billion for Analytics 360), GA4 may use a sample. The data quality icon then shows the share of data used.
- Thresholds: with demographics, search terms or very small numbers, GA4 withholds values so nobody can infer individuals. A longer date range often helps.
For small websites the second point matters more: if a report looks empty or incomplete, it’s often thresholds, not lack of interest.
Consent gaps: what’s missing from your numbers
In the EU, analytics with cookies only loads after consent. People who decline are missing from classic numbers. Google offers Consent Mode for this: when consent is denied, tags send cookieless pings, per Google, from which GA4 estimates the gaps by modelling.
- Modelling needs the advanced implementation, where Google tags load even without consent; if you block tags until consent, there’s no modelling.
- If consent is managed via the IAB TCF, GA4 can’t model data for users who decline.
- Modelled values are estimates, not measurements.
Whether and how you use Consent Mode is a decision for your data protection advice. For analysis: expect your numbers to miss some visitors or include them only as estimates, and compare trends rather than absolute values.
A robust complement are metrics that need no tracking: enquiries in the inbox, shop orders, newsletter sign-ups, clicks and impressions in Search Console.
Search Console and Looker Studio
Google Search Console shows which search queries your pages appear and get clicked for in Google Search. It needs no cookies on your site, so it’s a good complement to GA4. Per Google, the newest data is preliminary and may change over the next few hours.
- Usable with AI: export the performance report (queries, pages, clicks, impressions, position) and have AI find patterns: pages with many impressions and few clicks, queries without a matching page.
- Looker Studio: the free version combines GA4, Search Console and spreadsheets into one dashboard, so you see the key numbers in one place instead of pulling them from five tools every week.
How to improve content with this data is covered in AI for SEO; how to become visible in AI answers and measure it, in GEO in practice.
AI analysis tools for your own spreadsheets
For exports from ad accounts, shops or newsletter tools, use tools that read spreadsheets and show their calculation steps. The difference from a general chat: they run code instead of estimating numbers.
| Tool | Best for | More |
|---|---|---|
| Julius AI | Analyse Excel, CSV and database tables by asking, with charts and visible calculation steps | Overview · Try →Ad |
| Looker Studio | Free dashboards from GA4, Search Console and spreadsheets, no partner link | |
| ChatGPT, Claude | With file upload and code execution for simple analyses, no partner link |
“Try” links are affiliate links (advertising): I earn a commission at no extra cost to you.
| Tool | Price per vendor (Nov 15, 2026) | Source |
|---|---|---|
| Julius Plus | $20 a month | julius.ai |
| Julius Pro | $45 a month | julius.ai |
| Julius Business | $450 a month, team up to 50, database connectors | julius.ai |
| Looker Studio | Free version available | Google Cloud |
Per Julius’ pricing page, annual billing saves 20%.
Whatever the tool: ask for the calculation path. A result without traceable steps isn’t good enough for decisions.
Which metrics really count
Many dashboards show what’s easy to measure, not what matters. Choose metrics by your business goal.
| Goal | Core metric | Leading indicator |
|---|---|---|
| Enquiries for a service | Qualified enquiries per week | Visits to offer pages |
| Online shop | Revenue and contribution margin per channel | Cart abandonment, returning visitors |
| Growing a newsletter | Confirmed sign-ups | Visits to pages with a sign-up form |
| Visibility | Clicks from Google Search, mentions in AI answers | Impressions in Search Console |
Page views, time on page and followers are helper metrics. They explain why the core metric moves but aren’t goals in themselves.
UTM parameters: the basis for clean attribution
Without clean tagging, visits from newsletters, social media and ads end up as “direct” or “unknown”. UTM parameters on the link fix that.
- utm_source: where from (newsletter, instagram, linkedin).
- utm_medium: what kind (email, social, paid).
- utm_campaign: which campaign (black-friday-2026).
- utm_content: which variant (image-a, copy-b), important for tests.
Keep spelling consistent, ideally in a small table AI generates from a campaign list. Otherwise “Newsletter” and “newsletter” become two rows in GA4.
Reading AI reports in ad accounts critically
Meta and Google grade their own campaigns. Their AI recommendations aim to get you to put more budget into their automation. That can be right, but it isn’t neutral advice.
- Compare reported results with your own data: enquiries in the inbox, orders in the shop.
- Look at the breakdown by placement and device before accepting recommendations.
- Watch the attribution window: what the platform counts as a result depends on how long after a click or view it credits results to itself.
More in AI in advertising.
Measuring whether AI assistants name you
A growing share of research happens in ChatGPT, Claude, Gemini or Perplexity. Those conversations don’t show up in any analytics software. At most you see visits that come from a link in an AI answer.
- Watch GA4 and your server logs for referrals from AI services and fetches by their crawlers.
- Regularly ask real customer questions in several assistants and note who gets named.
- Ask new customers in your form or conversation how they found you; “via ChatGPT” is an answer no tool measures.
To start, our AI duel shows whom AI assistants recommend for typical customer questions, and the AI visibility check tests how readable your website is for AI.
Check data quality before AI analyses it
AI analyses what you give it. Five checks before any larger analysis prevent most wrong conclusions:
- Gaps over time: were there days without data, e.g. after a website change?
- Double counting: does an event fire twice, e.g. through two tags?
- Own visits: are tests and internal visits filtered out?
- Source spelling: do the same channels appear under different names?
- Definitions: does “conversion” mean the same in every tool?
The honest limits: made-up numbers and wrong conclusions
- Made-up numbers: language models without a calculation tool estimate numbers and sound confident doing it. Always ask for code execution or check the maths.
- Wrong period or filter: AI summaries don’t always state which date range and filters apply. Check both.
- Correlation, not cause: more revenue after a campaign doesn’t mean the campaign caused it. Season, prices and other channels play a part.
- Small numbers: 12 sign-ups instead of 8 is 50% more, but often chance.
- Bot traffic and your own visits: automated fetches, check scripts and your own tests distort visit numbers if not filtered out.
From practice: how we measure
For i4u.business we count visits without cookies from the server logs, filter out bots and our own visits, and add Search Console. Ad tracking only runs after consent.
The key lesson: our own check scripts, Lighthouse runs and tests can distort the stats more than outside bots. Since we tag them with our own identifier and exclude them, the numbers are much more honest. And we compare trends over weeks, not single days.
A 30-minute weekly analysis routine
- Get the numbers (10 min): five metrics from GA4, Search Console, newsletter tool and inbox into one table.
- Ask AI (5 min): what stands out compared with the last four weeks? Show the calculation.
- Check (10 min): look up the two most striking points in the original report.
- Decide (5 min): one thing you’ll do differently next week, and how you’ll measure whether it helped.
How to generate such reports automatically is shown in marketing automation with AI.
Checklist for AI in analytics
- Tracking plan: which events, which goals, which sources
- Consent setup checked, gaps known
- Search Console set up and linked to GA4
- Own visits and check scripts filtered out
- Five core metrics set, including some without tracking (enquiries, orders)
- AI analyses only with a visible calculation path
- Date range and filters checked for every AI statement
- Trends over weeks rather than single days
- Weekly 30-minute routine in the calendar
12 prompts for marketing analytics
Create a tracking plan for [site type] with goals [goals]: which events, which parameters, which reports. Short and actionable.
Here are the numbers for the last eight weeks: [table]. Which values deviate notably? Calculate with code and show the steps.
Summarise this report in five sentences for management. State the date range and filters explicitly: [report].
Revenue rose in week [X]. What are possible explanations (season, campaign, price, channel)? What data would I need to tell them apart?
Here’s a Search Console export: [data]. Which pages have many impressions but few clicks? Which queries have no matching page?
Compare these campaigns by cost per result and result quality, not clicks: [data]. Calculate with code.
Here’s the breakdown by placement: [data]. Where do lots of cheap clicks without results come from?
Variant A: [values], variant B: [values]. Is the difference at this volume probably chance? Explain simply.
Plan a Looker Studio dashboard with at most eight metrics for [goal] from GA4, Search Console and a spreadsheet. Which charts, which filters?
Our consent rate is about [value]. Which reports are most affected, and which metrics are independent of it?
Write a weekly report from these five metrics with three anomalies and one recommendation, max 150 words: [data].
Here’s an AI analysis: [text]. Which numbers should I verify in the original, and which conclusions are too strong?
The most common mistakes
- Letting a chat estimate numbers. Always have it calculate with code, or check yourself.
- Reading empty reports as lack of interest. Often it’s thresholds or consent gaps.
- Optimising clicks instead of results. Cheap clicks without results only cost money.
- Counting your own visits. Tests and check scripts distort the stats.
- Comparing single days. Trends over weeks are more reliable.
- Believing an AI overview without the report. Check the original first, then act.
Frequently asked questions
Is Analytics Advisor available in German?
Not yet, per Google Help on November 15, 2026: the assistant is currently limited to properties set to English; Google says more languages are planned.
Can I do marketing analytics without Google Analytics?
Yes, within limits. Server logs, Search Console, newsletter tool, shop and CRM data provide many metrics without cookies. For on-site behaviour analysis you need an analytics tool, with consent where required.
How do I present numbers to management?
With a few core metrics, a comparison with the previous period, a clear statement of what changed and a recommendation. AI can draft it; you check date range, filters and numbers yourself.
How do I measure whether ChatGPT recommends my company?
Analytics tools only show visits via links in AI answers. Add regular spot checks with real customer questions in several assistants, and ask new customers how they found you.
What should I check before analysing data with AI?
Gaps over time, double-counted events, your own visits, inconsistent source names, and whether terms like “conversion” mean the same everywhere.
Which marketing metrics should I measure?
The ones that fit your business goal: e.g. qualified enquiries, revenue per channel or confirmed newsletter sign-ups. Page views and followers are helper metrics that explain why the core metric moves.
Why do I need UTM parameters?
So visits from newsletters, social media and ads are attributed to their source in analytics instead of showing up as “direct”. Consistent spelling matters.
Can I analyse GA4 with AI?
Yes. GA4 shows AI overviews of important changes and is testing Analytics Advisor, an assistant built on Gemini, currently only for properties set to English. Always check statements in the linked report.
Why are some GA4 reports empty or incomplete?
Often because of thresholds: GA4 withholds data when too few users are in a report, e.g. with demographics or search terms. A longer date range helps. On large queries GA4 may also use sampling.
Does ChatGPT make up numbers?
Language models without a calculation tool can estimate numbers and sound confident. For analyses, use tools or modes that run code and show their steps, and check key values yourself.
What does Julius cost?
Per its pricing page on November 15, 2026: Plus $20, Pro $45 and Business $450 a month; annual billing saves 20% per the vendor.
Do I need Looker Studio?
Not necessarily. It pays off if you want to see numbers from several sources regularly in one place. The free version is enough for most small teams.
How do I handle data missing because of declined cookies?
Compare trends rather than absolute values, and add metrics that need no tracking, such as enquiries, orders, newsletter sign-ups and Search Console data. Whether Consent Mode suits you is a question for your data protection advice.
How often should I look at marketing numbers?
Once a week in a fixed routine is enough for most. Checking daily tempts you to react to random swings.
Sources
- Google Analytics: Generated insights – AI overviews
- Google Analytics: Analytics Advisor (beta) – Gemini, English only, may make mistakes
- Google Analytics: Data thresholds – Thresholds
- Google Analytics: Data sampling – 10M / 1B events
- Google Analytics: Consent mode – Modelling, requirements
- Search Console: Performance report – Newest data preliminary
- Google Cloud: Looker Studio – Free version
- Julius: Pricing – Plus $20, Pro $45, Business $450






