The Short Version
You can now measure part of your AI search visibility with Google's own data. Since 31 August 2026, Search Console's Generative AI performance report has been available to all websites. It shows impressions in AI Overviews and AI Mode, by page, country, device and date.
It does not show clicks, and it says nothing about ChatGPT, Perplexity or Copilot. For those, you combine GA4 referral data with a fixed list of prompts you test every month. Set up all three once and the monthly check takes under an hour.
Most owners I work with now ask some version of the same question: how do we measure AI search visibility, and is it worth paying attention to yet? It usually comes after someone asked ChatGPT about their industry and saw a competitor named instead. Until recently, though, the honest answer was that nobody could measure it properly, including the tools selling AI visibility dashboards.
That changed a little this year. Google added AI impressions to Search Console, and the rollout reached every website at the end of August. It's not the full picture, but it's real data from Google rather than an estimate. Below is the setup I now use for clients in Malaysia: three layers, and what each can and can't tell you.
🧠Context
Why This Question Keeps Coming Up
AI answers change what a "ranking" means. In classic search you either appear in the ten blue links or you don't. In an AI Overview, Google writes an answer and shows a handful of supporting links beside it. In ChatGPT, a brand may be named in the answer with no link at all. Being mentioned and being clicked are now different things.
If you've read my piece on what AEO and GEO mean for a business, you'll know I don't think every SME needs to rebuild its marketing around AI. But you should know roughly where you stand. Otherwise you end up making decisions on a screenshot someone forwarded in a group chat.
So the goal of measurement here is modest. You want to answer three questions every month:
- Does Google show our pages inside AI Overviews and AI Mode, and which pages?
- Are AI assistants sending us any visitors, and do those visitors do anything useful?
- When a buyer asks an assistant about our service in Malaysia, are we named, and is the description accurate?
Each question needs a different source.
📈Layer 1
Layer 1: The Search Console Generative AI Report
Google announced the Search Generative AI performance reports in June 2026 and first released them to a subset of sites. According to Google's help page for the report, they were rolled out to all websites worldwide as of 31 August 2026. If you haven't opened Search Console since then, it's worth a look.
What it shows
The report counts impressions: how many times links to your site were shown to a user in a generative AI feature on Google Search. It covers AI Overviews and AI Mode. Search Labs experiments are excluded, and Discover has a separate report of its own.
You can break the data down four ways, which will feel familiar if you already use the normal Performance report. My Search Console guide for Malaysian SMEs covers that report if you need a refresher.
| Dimension | What it tells you | What I use it for |
|---|
| Pages | Which of your URLs appeared in AI features (canonical URL, after redirects) | Finding the pages Google trusts enough to cite |
| Countries | Where the searches came from | Filtering to Malaysia before reading anything else |
| Dates | Days, weeks or months, in Pacific Time | Spotting changes after a content update or a Google update |
| Devices | Desktop, mobile or tablet | Checking whether AI visibility is mainly a mobile story for you |
There's also a search type filter that separates normal text searches from multimodal ones, meaning image based searches such as Google Lens and Circle to Search. For most B2B firms, the text figure is the one that matters.
What it does not show
This is where people get caught out. The help page describes impressions only. There's no click column and no position. So you can see that a page was shown in an AI Overview 400 times last month, but not how many of those people clicked through.
A few other details matter when you read the totals:
- In the chart, two results from your site in the same AI feature count as one impression. Table totals are aggregated differently, so the chart and table won't always match.
- The usual Search Console limits apply, including the 1,000 row limit in the interface.
- The newest days can be preliminary and may change.
- If your site has very few AI impressions, the report may not appear at all. That's information too.
One more point that confuses clients. AI Overview and AI Mode traffic was already included in your normal Performance report under the Web search type. The new report doesn't add traffic you weren't counting before. It separates out the AI part so you can see it on its own.
💡
The comparison worth making
Export the Pages table from the AI report and the Pages table from the normal Performance report for the same 28 days, filtered to Malaysia. Pages with high AI impressions but few normal clicks are often answering the question so well that people don't need to click. Pages with no AI impressions at all are the ones to look at first.
📊Layer 2
Layer 2: AI Referral Traffic in GA4
Search Console only covers Google. For visitors who come from ChatGPT, Perplexity, Gemini or Copilot, GA4 is the place to look, because these tools usually pass a referrer when someone clicks a link in an answer.
Find what you already have
Open Reports, then Acquisition, then Traffic acquisition. Change the primary dimension to Session source and search for chatgpt, perplexity, gemini and copilot. On most SME sites I audit, the numbers are small. A few dozen sessions a month is common for a professional services firm.
Make it permanent with a custom channel group
Rather than searching every month, set up a custom channel group in GA4 Admin with a channel called AI assistants. Use a rule where Session source matches a regular expression such as:
chatgpt|openai|perplexity|gemini|copilot|claude
Place it above the Referral channel so these visits don't get lumped in with ordinary referrals. GA4 applies custom channel groups to past data as well, so you'll see history straight away.
The caveats
This layer undercounts, and you should expect it to. Some AI app visits arrive with no referrer and show up as Direct. People who read an answer, then type your brand into Google, show up as organic search. And it says nothing about the many answers where you were mentioned without a link. Treat GA4 as a floor, not a total.
What GA4 does well is quality. Compare engaged sessions and key events for the AI assistants channel against organic search. Small volumes can still be worth watching if those visitors enquire.
🔍Layer 3
Layer 3: A Monthly Prompt Test
The third question, whether you're named when buyers ask, can only be answered by asking. It's manual, and it's also the part that tells you the most.
Build a fixed prompt list
Write 20 to 30 prompts that a real buyer might type. Keep the list fixed so month to month changes mean something. I split them into four groups:
| Prompt type | Example | How many |
|---|
| Category | Who are good ISO consultants in Selangor? | 6 to 8 |
| Problem | How do I get ISO 9001 certification for a small factory in Malaysia? | 6 to 8 |
| Comparison | Should I hire a consultant or do ISO certification myself? | 4 to 6 |
| Brand | What does [your company] do, and is it reliable? | 3 to 4 |
Include a few in Bahasa Malaysia if your buyers search that way. In my experience the answers can differ quite a bit between English and BM, sometimes naming completely different firms, because the sources the assistant draws on are different.
Run it the same way each time
- Use the same tools each month. For most Malaysian SMEs that's Google AI Mode, ChatGPT and one of Gemini or Perplexity.
- Use a logged out or fresh session where you can. Personal history affects answers.
- Log four things per prompt: were you named, which competitors were named, which of your pages was linked, and was anything said about you wrong.
- Run each important prompt twice. AI answers vary between runs. If you're named once out of two runs, record it as partial.
A shared spreadsheet is enough. If you want to act on what you find, my earlier post on how to rank better in ChatGPT covers the content side.
⚠️
Watch for wrong answers, not just missing ones
The most useful finding from prompt tests is often an error: an old address, a service you stopped offering, a wrong price range. Those usually trace back to an outdated page on your own site or a stale directory listing, and they're fixable.
🧩Analysis
How to Read the Numbers Together
Each layer on its own is easy to misread. Together they usually point to one clear next step. These are the patterns I see most often.
| What you see | What it usually means | What to do |
|---|
| AI impressions on blog posts, none on service pages | Google sees your guides as useful, but your commercial pages don't answer questions directly | Add plain answers, FAQs and specifics to service pages |
| Almost no AI impressions at all | Often an indexing or snippet problem rather than an AI problem | Check indexing and any nosnippet or max-snippet settings first |
| Named in ChatGPT, absent in AI Mode | Your brand is known from other sources, but your pages don't rank well in Google | Work on classic SEO for those pages |
| AI referrals with high engagement | Small but qualified traffic | Find the landing pages and build on those topics |
| Prompt tests show wrong facts | Outdated pages or listings being quoted | Fix the source page, then retest next month |
The second row deserves a note. Google's own documentation on AI features says a page must be indexed and eligible to be shown with a snippet to appear as a supporting link, and that there are no additional technical requirements. So before anyone sells you an AI visibility package, rule out the basics. I check this as part of every SEO audit, and it's surprising how often a stray snippet setting or a batch of unindexed pages explains the gap.
The same page says you don't need special schema or AI text files to appear in these features, so nobody should tell you a file is the missing piece.
If the problem turns out to be structural, the sort of thing covered in our technical SEO checklist, it's worth fixing before you write anything new. And if you're planning a rebuild soon, read my notes on a website redesign SEO checklist first, because a careless relaunch is one of the fastest ways to lose both classic and AI visibility at once.
✂️Trade offs
What Is Not Worth Measuring
There's a growing market for AI visibility scores, share of voice percentages and daily tracking across every assistant. Some are useful for large brands. For a typical Malaysian SME, I'd hold off, for three reasons.
- The sample sizes are small. A score based on a few hundred automated prompts can swing a lot from week to week without anything real changing.
- Daily tracking invites overreaction. AI answers vary run to run. Checking monthly gives you a steadier picture and stops the team rewriting pages every time an answer changes.
- The decisions are the same. Whatever the score, the work is usually the same: clearer pages, better answers to real questions, accurate business details everywhere, and solid classic SEO.
I'd also skip tracking AI impressions for queries you don't sell to. Focus on the pages and prompts tied to what you actually offer, in the cities you serve.
If you'd rather have someone run this for you, it's part of our AI search optimisation work. But the setup above is designed so you can do it in house.
What to Set Up This Week
Open Search Console and check whether the Generative AI performance report appears for your property. If it does, filter to Malaysia, export the Pages table, and save it as your baseline. Then create the AI assistants channel group in GA4 so the history starts collecting in one place.
Finally, write your prompt list and run it once. It will take about an hour the first time. After three months you'll have real data on all three questions, and you can decide whether AI search deserves more of your budget or whether, for now, technical SEO and better service pages are the better spend. For many firms it's the second.
Frequently Asked Questions
Does Google Search Console show AI Overview traffic?
Yes. Since 31 August 2026 the Generative AI performance report is available to all websites. It shows impressions in AI Overviews and AI Mode by page, country, device and date. It does not show clicks or position, and AI traffic is also still counted in the normal Performance report under Web.
Can I see ChatGPT traffic in Google Analytics 4?
Partly. When someone clicks a link in a ChatGPT answer, the visit usually appears in GA4 with chatgpt.com as the session source. Some visits arrive with no referrer and are counted as Direct, so treat the GA4 figure as a minimum rather than the full number of AI visits.
How often should I measure AI search visibility?
Monthly is enough for most Malaysian SMEs. AI answers vary between runs, so daily checks mostly show noise. A monthly routine covering the Search Console AI report, the GA4 AI channel and a fixed prompt test takes about an hour once it is set up.
Why does my site have no data in the AI performance report?
Usually because the site has too few impressions in AI features for the report to appear. It can also mean pages are not indexed or have snippet restrictions such as nosnippet. Check indexing in the Pages report before assuming it is an AI specific problem.
Do I need special schema or an AI text file to appear in AI Overviews?
No. Google's documentation says there are no additional requirements or special optimisations needed for AI Overviews or AI Mode, and no special schema or AI text files. A page needs to be indexed and eligible to show with a snippet, the same as normal search.
Should I test prompts in Bahasa Malaysia as well as English?
If your customers search in Bahasa Malaysia, yes. Answers can differ between languages because the assistant draws on different sources. Add a few BM versions of your most important prompts to the monthly list and log the results separately so you can compare them.
