Is your AEO agency working? What the report must show

A working AEO program shows up as movement in a fixed set of buyer questions, asked repeatedly across several AI engines and recorded over time. One screenshot proves nothing, because about 17% of prompts return a different set of recommended brands than they did the day before (MaxAEO, 2026). A real monthly report names the prompts, the engines and the baseline.

Key takeaways

• AI answers move on their own. MaxAEO diffed 897,840 answers over 90 days and found "about 17% of prompts return a different set of recommended brands than they did the day before" (MaxAEO, 2026).
• Search metrics do not stand in for AI visibility. Across 1,094 US ChatGPT categories, Semrush found only 15.2% had a clear owner, and owners had higher organic traffic in just 48.4% of paired comparisons (Semrush, 2026).
• A small traffic number is normal. Conductor measured AI referral traffic at 1.08% of all website traffic across the ten industries it tracks (Conductor, 2026).
• Your analytics undercounts it anyway. Of 371,847 sessions measured in April 2026, "35.7% of AI traffic arrived with no referrer at all", landing those visits in Direct (Clickport, 2026).

New to this? Acclaira runs a free one-hour AEO class on Zoom, every 2 weeks on a Thursday at 9.30pm Malaysia time. Ask for the Zoom link on WhatsApp.

What should an AEO monthly report actually contain?

An AEO monthly report should contain five things: the fixed prompt set you agreed on, the engines each prompt ran on, the date of every run, your result in each run recorded as cited, mentioned or absent, and the change since the baseline. Traffic and enquiries sit around that record as context, not in place of it.

The prompt set is the part most clients never see, and the part that decides whether the rest of the report means anything.

Around that spine sit the enquiries that mentioned AI, tagged at intake, and a record of what actually changed on the site. This is the standard Acclaira holds its own client reporting to; our DNE Logistics case study is the worked before-and-after on a real Malaysian business.

Why is one screenshot of ChatGPT naming you not proof?

Because AI answers are unstable by design, so a single screenshot captures one draw from a moving distribution. MaxAEO re-ran the same 1,247 buyer-intent prompts daily for 90 days across eight platforms and diffed 897,840 answers against the previous day's. About 17% of prompts returned a different brand set each day, and the median brand list survived unchanged for five days (MaxAEO, 2026).

That instability cuts both ways: a screenshot naming you is not proof the program worked, and one answer leaving you out is not proof it failed. Only repetition separates signal from noise.

Winning one question is not the same as owning a subject. Semrush's study of 1,094 US ChatGPT categories put a number on it:

"Winning one prompt in ChatGPT is not the same as owning a topic. Real topic ownership requires a brand to appear across at least four of five related prompts, with a 5-percentage-point lead over the runner-up."

Margarita Loktionova, Semrush, AI visibility is a topic-level game

The same study found domain-level search metrics "correlate with ownership only about half the time". A report handing you keyword rankings and calling it AI visibility is measuring a different thing, which is why a business can sit number 1 on Google and still be absent from ChatGPT.

Why do the AI traffic numbers look so small?

Because AI referral numbers are small even on large enterprise sites, and analytics loses a large share of them before they are counted. Conductor measured AI referral traffic at 1.08% of all website traffic across the ten industries it tracks, with ChatGPT accounting for 87.4% of it (Conductor, 2026). A Malaysian SME seeing a low double-digit session count is not seeing a failure.

The undercount is structural, not a setup error. Clickport measured 371,847 sessions in April 2026 and found 35.7% of AI traffic arrived with no referrer at all, landing in Direct (Clickport, 2026). Any AI session figure in your report is a floor, and an agency presenting it as a total either does not know that or is hoping you do not.

Small does not mean worthless. Patrick Stox reported that AI search sent Ahrefs 0.5% of its traffic but 12.1% of its signups, while warning that AI search users "click links 75% less than they do in traditional organic search" (Ahrefs, 2025). That is one company's own 2025 analytics, a reason to tag enquiries at intake, not a benchmark for your agency.

What are the red flags in an AEO report?

The clearest red flag is a report that leads with traffic: it is the easiest number to grow by other means and the hardest to attribute to AI. An honest report makes itself checkable. Seven warning signs are worth naming.

1. Traffic in the headline position, with the prompt-by-prompt record buried.
2. Screenshots with no date, no prompt text and no engine named.
3. A prompt set that quietly changes between months, making improvement unreadable.
4. A single blended "AI visibility score" with no published method behind it.
5. Any guaranteed ranking, guaranteed citation or promised date. Nobody controls the engines.
6. No named competitors, so you never learn whether you gained or the category moved.
7. No record of what was actually changed on your site that month.

Our fuller guide to choosing an AEO agency in Malaysia covers the pre-hire stage and keeps its own six-point version of that checklist; what AEO should cost in Malaysia sets out what the fee buys. Acclaira's own program is RM2,500 a month, billed from day one, month to month. This standard binds us too.

How do I check the report myself?

Take three prompts straight out of the report, open ChatGPT, Gemini and Perplexity with web search switched on, and ask them exactly as written. Do it twice, two days apart, and save both with the date visible. Drift between your runs and the agency's is expected. A result that never reproduces at all is not.

This costs nothing: you can do this much yourself, without hiring anyone. What it does not replace is a fixed prompt set run to a schedule against a baseline, month after month.

Check whether the answer links to your site or only names you. A mention builds recognition, a citation sends a visitor, and a report merging the two hides the weaker half. Our guide to checking whether ChatGPT recommends your business sets out the full prompts and scoring.

Why does AEO mean something else in Malaysia?

In Malaysia the acronym belongs to someone else, so write "Answer Engine Optimization" in full when you test. On 7 August 2026, Acclaira asked ChatGPT, Gemini and Claude, with web search enabled, how to tell whether an AEO agency is working.

ChatGPT answered entirely about Authorized Economic Operator customs accreditation. Claude stopped to ask which AEO was meant, pointing at the Royal Malaysian Customs Department programme. Only Gemini read it as Answer Engine Optimization.

The lesson is bigger than the acronym: your prompt set should use the words a customer would actually say, not industry jargon.

Frequently asked questions

How soon should an AEO report show movement?

Treat month one as the baseline and expect the first readable trend around month two or three. The pace depends on how often your pages are re-indexed, your site's depth and your review activity. No honest agency will guarantee a date, because none of them control the engines.

Should my AEO report include keyword rankings?

As context, yes. As the headline, no. Semrush found domain-level search metrics correlated with ChatGPT category ownership only about half the time, with organic traffic matching in 48.4% of paired comparisons (Semrush, 2026).

My AI referral sessions are tiny. Is the program failing?

Not on that evidence alone. Conductor put AI referral traffic at 1.08% of all website traffic across the ten industries it tracks (Conductor, 2026), and 35.7% of AI visits arrive with no referrer (Clickport, 2026). Judge the program on cited-versus-absent counts and on enquiries.

The report shows a competitor and not us. Is that a bad report?

That is a good report. A record showing you absent while naming who was cited instead is the only version that tells you what to fix next month.

Want the method walked through with your own business as the example? Acclaira's free one-hour AEO class runs on Zoom, every 2 weeks on a Thursday at 9.30pm Malaysia time. Message us on WhatsApp to join.

Sources

• Chris Han, "How Often Do AI Answers Change? 90-Day Data From 8 Platforms", MaxAEO, 11 June 2026 (1,247 buyer-intent prompts re-run daily for 90 days across eight AI platforms; 897,840 answers diffed against the previous day's; study window 1 March to 29 May 2026)
• Margarita Loktionova, "AI visibility is a topic-level game: A study of 50,000 brands in ChatGPT [Study]", Semrush, 20 July 2026 (1,094 US ChatGPT categories, over 50,000 brands, January to June 2026)
• Conor Baker and Wesam Rafidi, "The 2026 AEO / GEO Benchmarks Report", Conductor, last updated 6 July 2026 (AI referral traffic across 10 key industries; wider analysis covers 1,215 enterprise customer domains and more than 3.3 billion sessions)
• David Karpik, "Why ChatGPT Traffic Shows as Direct in GA4 (2026)", Clickport, last updated 21 May 2026 (371,847 sessions measured, April 2026)
• Patrick Stox, "Does AI Search Traffic Convert Better Than Traditional Search? For Ahrefs, Yes: 0.5% of Visitors Drove 12.1% of Signups", Ahrefs, 16 June 2025 (single-company data from Ahrefs Web Analytics)
• Acclaira first-party engine probe, 7 August 2026: three buyer questions asked of ChatGPT, Gemini and Claude with web search enabled. Per-query answers and cited-source domains retained in the run's probe JSON.
• Acclaira free AEO class (every 2 weeks on a Thursday, 9.30pm Malaysia time, on Zoom, free)