Ten Malaysian agencies sell AI search optimisation. We measured all ten on four checks, put ourselves in the table, and then asked four AI engines to recommend an agency.
We came top of the technical checks. Not one of the four engines named us. The agency that scores zero on every check was named by three of them.
That result is the reason this page exists. If you are about to pay someone to get your business mentioned by ChatGPT, the gap between those two facts is the most useful thing you can know.
Everything below was measured on 18 August 2026 and every check is one you can run yourself in about five minutes. Where we could not get a reading, the cell says so rather than staying blank.
The short version {#sec-0}
If you only read one section, read this one.
- Ten agencies, four checks, one afternoon. Crawler access,
llms.txt, Organization schema, FAQPage schema. No paid tools needed. - Only three of ten name AI crawlers explicitly. Seven leave it to a catch-all rule.
- The technical checks do not predict who gets recommended. We top them and no engine names us. LOCUS-T scores zero and three engines name it.
- Each engine reads something different, and the source changes with the question. We asked Perplexity two questions on the same day and got directories once, agencies' own service pages the other time.
- AI sent 7 visits out of 1,707 across five sites we run. Google sent 502 over the same window. And the biggest batch inside that 7 does not survive scrutiny.
- The two agencies whose prices got quoted back by an AI were the two who publish prices on their own pages.
How we measured this {#sec-method}
Four checks, run the same way on every domain including our own, on 18 August 2026.
We fetched each site's robots.txt and counted every crawler named on it, then cross-checked that list against the same set of AI user agents for all ten. We fetched /llms.txt and treated an HTML error page as absent. We pulled each homepage and parsed its JSON-LD for Organization and FAQPage types.
Two things about method that changed the results:
We counted our own row the same way we counted everyone else's. The first version of the table used a script with an incomplete crawler list, and the incomplete list happened to flatter us. Recounting moved Atlas Studio from 7 to 10 and erased our lead.
One domain returned an error and we left the row blank. Writing "0" would have been easier and would have made us look better against it.
The four checks, and what they actually mean {#sec-1}
None of these require a paid tool. They test whether a site has done the basic work of being readable by machines.
| Check | How you run it | What it means |
|---|---|---|
| AI crawlers named | Add /robots.txt to their domain |
Whether GPTBot, ClaudeBot, PerplexityBot and friends are explicitly allowed by name rather than left to a catch-all |
| llms.txt | Add /llms.txt to their domain |
A plain-text summary of the business written for AI clients that read it |
| Organization schema | Paste their homepage into Google's Rich Results Test | Machine-readable facts about who the company is |
| FAQPage schema | Same test | Question-and-answer pairs an engine can lift directly |
Does an llms.txt actually do anything?
A warning about the second check, because we sell this work and it would be easy to oversell it. Google's own documentation states plainly that Google Search ignores llms.txt and that keeping one "will neither harm nor help your site's visibility or rankings." It matters only for AI clients that do read it. We keep ours for that reason and we tell clients the difference. More on that in our breakdown of what GEO actually is.
The results {#sec-2}
| Agency | AI crawlers named | llms.txt | Organization schema | FAQPage |
|---|---|---|---|---|
| FoundlyAgency (us) | 11 | ✅ | ✅ | ✅ |
| Atlas Studio Digital | 10 | ✅ | ✅ | ✅ |
| MediaPlus Digital | 3 | ✅ | ✅ | ✅ |
| Cybergate | 0 | ✅ | ✅ | ✅ |
| ZenWeb | 0 | ✅ | ✅ | — |
| Xantec | 0 | — | ✅ | ✅ |
| SEO.my | 0 | — | ✅ | — |
| LOCUS-T | 0 | ✅ | — | — |
| VeecoTech | 0 | ✅ | — | — |
| Nexus Mediaworks | not measured | not measured | not measured | not measured |
Three of ten, and seven that leave it to a catch-all
Three of the ten name AI crawlers explicitly. Seven carry an llms.txt.
Why our own row is a tie, not a lead
Our 11 versus Atlas Studio's 10 is a tie, not a lead — and their list is arguably better than ours. They name Google-Extended, Bytespider, cohere-ai and Meta-ExternalAgent. We name none of those four. We cover more of the OpenAI and Anthropic variants; they cover more vendors. If you are choosing between us on this row, it is a wash.
We mention that because the first version of this table said 7 for Atlas Studio. That number came from a script with an incomplete list of crawler names, and the incomplete list happened to flatter us. Counting every crawler on every robots.txt the same way moved them from 7 to 10 and closed the gap. A comparison table is only worth reading if the author counted their own row the same way they counted everyone else's.
The row we left blank
Nexus Mediaworks returned an HTTP 530 error when we checked, so we have no reading. That row is blank because our measurement failed, not because the agency did nothing. Writing "0" there would have been the easier and more flattering choice, and it would have been false.
Now the part that undermines the table {#sec-3}
On the same day, we asked four AI engines to recommend a Malaysian web and SEO agency, using the kind of question a business owner would actually type.
None of the four named us — the agency at the top of that table.
Here is the same table with one column added.
| Agency | AI crawlers named | Named by an AI engine? |
|---|---|---|
| FoundlyAgency (us) | 11 | ❌ None of four |
| Atlas Studio Digital | 10 | ✅ ChatGPT, first pick |
| MediaPlus Digital | 3 | ✅ Google AI Mode |
| Cybergate | 0 | ✅ ChatGPT |
| ZenWeb | 0 | ✅ Google AI Mode, Gemini |
| LOCUS-T | 0 | ✅✅✅ Google, Perplexity and Gemini |
The agency that does none of it and gets named by three engines
LOCUS-T names no AI crawlers, carries no Organization schema, and no FAQPage. Three engines out of four recommend it. We do all of it and none do.
So the technical layer is not what decides whether an AI recommends you. It is table stakes — worth having, cheap to do, and not the lever. Anyone selling you schema and an llms.txt as the route to being recommended by ChatGPT is selling you the part that is easy to bill for.
What the engines are actually reading {#sec-4}
We read the citation badges on each answer to see where the recommendations came from. They do not come from the agencies' own technical setup. They come from whatever source the engine happened to pull.
| Engine | What it cited |
|---|---|
| Google AI Mode | Listicle articles — "Top 5 SEO Companies", "Top 21+ Web Design Agencies" |
| Perplexity | Directory category pages — Sortlist, TechBehemoths |
| ChatGPT | The agencies' own service pages |
| Gemini | No sources shown |
Ask a different question, get a different source
There is a trap in that table, and we walked into it ourselves before catching it.
Those sources change with the question. We asked Perplexity two different questions on the same day. Asked for web design agencies in Petaling Jaya, it cited directories. Asked how a business gets recommended by ChatGPT, it cited eight agencies' own service pages and one trade publication — no directories at all.
So "Perplexity reads directories" is not a fact about Perplexity. It is a fact about one question we asked. If an agency tells you they know how a given engine picks sources, ask which query they tested and when.
The only two prices an AI quoted back
One thing did hold across both: the two agencies whose prices got quoted back inside the answer were the two who publish prices on their own pages — Xantec at "from RM6,000/month" and AI SEO Malaysia at "from RM3,500/month". Publishing a specific, checkable number appears to make it liftable. That is a cheap thing to act on.
How much traffic does any of this bring {#sec-5}
This is where most agency pages stop talking. Ours are the numbers from the five sites we run and maintain, over the ninety days to 18 August 2026.
| Site | Total sessions | From AI platforms |
|---|---|---|
| Our own site | 88 | 0 |
| Client A (supplements, Europe) | 432 | 5 |
| Client B (events) | 585 | 1 |
| Client C (restaurant) | 461 | 1 |
| Client D (retail) | 141 | 0 |
| Total | 1,707 | 7 — 0.41% |
Google sent 502 sessions over the same period. That is roughly 72 times more than every AI platform combined.
Why we do not trust our own best number
And the single largest batch in that column does not survive a closer look. Those five sessions on Client A all landed on the homepage rather than any specific page, all fell inside a three-day window, came from two countries that do not match the product, and produced a WhatsApp click rate of 80%. That is the fingerprint of someone testing their own site, not of customers arriving. Our honest read of our own data is that real AI-referred traffic across five sites is close to zero.
What this number cannot tell you
Two caveats, one in each direction. This counts clicks only — an engine naming your business without the reader clicking through never shows up in analytics at all, so 0.41% is a floor rather than the whole value. And these are five small sites over ninety days, which is a narrow sample.
We publish this because the alternative is quoting the impressive-sounding half. A page that showed you only the impressions and not the clicks would be telling you something true and leaving you with something false.
What this work costs in Malaysia {#sec-price}
Most agency pages in this market quote a range with no source. Below are the only figures we could read directly off a provider's own page on 18 August 2026, plus the one number we are missing ourselves.
| Provider | Published price | Where |
|---|---|---|
| AI SEO Malaysia | from RM3,500 / month | own site |
| Xantec | from RM6,000 / month | own site |
| Everyone else we checked | not published | — |
Why the quotes vary so much
Three reasons, and only the third one should worry you.
Scope. "AI SEO" covers anything from adding FAQ markup to a monthly programme with content production and digital PR. Two quotes ten times apart can both be honest.
Language count. A Malaysian programme running English, Bahasa Malaysia and Mandarin is three content operations, not one.
Relabelling. Some quotes are ordinary SEO with the acronym changed. The check is simple: ask what they will do that they were not already doing for SEO clients last year. If the answer is only "add schema and an llms.txt", read section 3 again.
The number we do not publish
Our own build prices are public and fixed, but AI search work sits inside our monthly plans rather than carrying a price of its own. We noticed this while writing this piece, and it is worth saying plainly: the two providers above got their prices quoted back inside an AI answer, and we did not. A specific, checkable number is easier for a machine to lift than a range.
What to ask before you hire anyone {#sec-6}
Copy these. They work on us too.
- "Show me your own baseline." Which exact prompts will you test in ChatGPT, Perplexity, Gemini and Google AI Overviews? What is my current mention rate, and which competitors come up instead of me? An agency that cannot show you this on its own site cannot build it for yours.
- "What happens if it is zero after sixty days?" A real answer names the fallback. No answer means nobody set a date to be judged on.
- "Which of these is table stakes and which is the lever?" If schema and llms.txt are described as the route to being recommended, compare that against the table in section 3.
- "How much traffic did AI send your last three clients?" Ask for the number with its denominator. Percentages without totals hide everything.
- Walk away from any guarantee of a number-one position in ChatGPT. AI answers vary by query, location, account history and model version. Nobody controls that, us included.
Where this leaves us {#sec-7}
We built the most machine-readable site of the ten we measured and it has not yet earned us a single AI recommendation. We are not going to pretend otherwise while selling the service.
What we will claim is narrower and checkable: we do the technical layer properly, and we measure the outcome. Every number on this page came from a script we run across all five sites, and we will re-run it and update this page with the date attached.
The next test is already scheduled. We will ask the same four engines the same questions on 17 October 2026, sixty days from publication, and update the table above with whatever comes back — including if the answer is still zero.
If you want the same measurement on your own site, that is the thing worth asking us for.
Method: robots.txt, llms.txt and homepage structured data fetched directly from each domain on 18 August 2026. AI engine answers collected the same day in a real browser. Traffic figures from Google Analytics 4 across five properties, 90 days to 18 August 2026. Nexus Mediaworks returned HTTP 530 and is recorded as not measured. Every check in section 1 is reproducible without a paid tool.