EcoWorld AI Marketing Masterclass · 28 July 2026

When buyers ask ChatGPT, will your property launch make the shortlist?

How to show up in AI answers. And the human-and-AI team that keeps you there.

Faisal Hourani

Buyer question

Which new launch near an MRT is best for a family with young children?

AI starting shortlist
Hypothetical example

Before we start

You have already done this yourself.

Think of the last big thing you bought. A car. A school. A holiday.

Most of us asked an AI first. And it changed what we looked at next.

The correction

Fixing this is a loop. Not a campaign.

Buyer asksWhich new launch near an MRT suits a family with young kids?
→
Find the gapWalk time never checked
→
Write the fixOne sourced answer
→
A person signs offApproves the claim
→
Ask againSame question, later
The facts move. So you ask again.

Write it once. Keep it true forever.

A couple with a toddler asks which new launch near an MRT would suit them. You go and ask the same thing yourself. Your project does not come up.

You dig. Every page you own says “steps from the MRT.” Nobody says how many minutes. And for this family it is not a commute question. It is a pram question. Eight minutes covered and flat is a different building from eight minutes across a main road in the sun.

So you write one line. Eight minutes on foot, measured on this date, and here is the route.

Someone puts their name on it before it goes live. If those eight minutes are wrong, that is not a typo. It is a claim about a home.

Then you ask again. And look where it ends. Right back at the start.

Part 1 · Look at the answer

Search hands you a list. AI hands you an answer.

Traditional search

family launch near MRT
1Developer launch page
2Property portal listing
3Map and neighbourhood guide
Pages to visit

AI answer

“Start with these three launches. One is closer to rail; another has larger family units…”

[1][2][3]
Shortlist + synthesis + sources

Simplified buyer-facing view. Engines and modes differ.

The hole in the middle

You can see the before and the after. You cannot see the answer.

You can see this They ask in sentences “family + MRT + new launch”
Search Console
→
You cannot see this The answer itself Who got named. What got cited.
What got said about you.
→
You can see this Someone turns up ChatGPT → your launch page
GA4
↑
Go and look yourselfAsk the question. Save the tool, the date, the answer and the sources. It is the only way in.

A search pattern is not a prompt log. Someone turning up is not a recommendation.

People used to type “condo KL.” Now they type whole sentences. That tells you they want an answer, not a list. And this is Google data. Nobody can see ChatGPT prompts. Not you. Not me.

Here is the hole. You cannot see the answer. Not who got named. Not what got cited. Not what got said about you.

Then somebody lands on your page from ChatGPT. That is real, and it is named. But it does not tell you how often you showed up, or what they asked.

So you go and look. That is the whole trick. There is no clever tool. You ask, and you write down what came back.

The test we actually run

So we go and test it. And we write down how.

  1. The questionFamily launch near rail
  2. Which toolOpenAI, with web search on
  3. The date20 July, fresh test
  4. What came backThree launches, plus links
  5. Check every claimRail, price, availability
A real OpenAI model response with web search to a Kuala Lumpur family property question, captured on 20 July 2026
OpenAI web-search response Captured 20 July 2026
What this audit foundCommercial sources named three launches. Verify route, date and availability before reuse.

One observation—not a permanent ranking.

Run this one yourself

Two minutes. Any assistant. Your own project.

Photograph this 
I'm looking at [YOUR PROJECT] in [AREA]
for a family with young children.

What can you tell me about it?

Then answer three more things:

1. Which sources did you use, and do any
   of them have something to sell?

2. What did you tell me that you could not
   verify from a primary source?

3. What would this family still need to know
   that you cannot answer?
Weak answer to 1?Your evidence lives on somebody else’s page.
Anything in 2?That is a wrong-fact risk. Go and fix the source.
Anything in 3?That is your first evidence gap. Start there Monday.

Turn web search on. Write down the date. One answer is an observation, not a ranking.

Four different wins

Getting named is not getting a visit. A visit is not a sale.

01Named?The launch shows up in the saved answer
→
02Came?A visit reaches the site from ChatGPT
→
03Did anything?Looked at units. Took the brochure.
→
04Enquired?A real enquiry from a real buyer

Do not squash four questions into one number.

Were you in the answer at all? That is the first gate. Most brands have never once checked.

Did anyone actually come? Different question. Different number. Different fix.

Did they do anything once they got here? Or did they look once and leave?

Did anyone enquire? That is the only one that pays for anything.

We changed the scoreboard

Our old scoreboard was missing how people found us.

Old view Search clicks One channel-shaped answer
→
New view Organic + named AI referrals Then engagement and business outcome

We changed the scoreboard before changing the strategy.

AI-referred2,106
Organic search1,423

On one of my own websites, more people arrived from AI than from Google search

Separate search context286,389 impressions · 1,204 clicksNot part of the GA4 comparison · 28 Mar–17 Jul 2026

An AI referral is an arrival signal—not proof of citation frequency, causality, or revenue.

Part 2 · Build the evidence

What actually happens when someone asks.

Splits the questionOne question becomes four or five
→
Goes and fetchesOnly what it can actually reach
→
Lifts out piecesNot pages. Pieces.
→
Writes one answerEvery sentence needs a source
→
They ask againAbout four turns, then a decision

Nobody outside the labs has the full recipe. But Google has published the shape, and it is enough to work with.

It does not search for what they typed. Google calls this fan-out. Their own example: “how do I fix a lawn full of weeds” becomes best herbicides, removing weeds without chemicals, how to stop them coming back. Three or four searches you never see.

Then it fetches. And it can only use what it can reach. Behind a filter, behind a login, inside a PDF, drawn by JavaScript. AI crawlers do not run your JavaScript. If the fact only appears after a click, the fact is not there.

It does not take your page. It lifts out pieces. And here is the part nobody expects: you draw the cut lines. Headings, tables, lists. Those are the seams. Google says the pieces do not need to be small. They need clear edges.

Then it writes. In the published research the model is told every sentence must carry a citation, and the source has to fully support that whole sentence. So a fact that needs three paragraphs of your page to make sense cannot be used at all.

Then they ask the next thing. Around four turns per conversation. Best areas. Then service charges. Then who actually delivers on time. Answer only turn one and you are gone by turn three.

Simplified. Engines differ and the exact recipes are not public. Sources: Google Search Central, and Aggarwal et al., “GEO”, KDD 2024.

The four checks

Take one fact. Ask four questions about it.

Photograph this
01 · AccessFind it

Can a buyer or a machine reach the fact?

The station, the route and the date are public.
02 · ExtractionUnderstand it

Does the fact make sense on its own?

The sentence stands up by itself.
03 · SupportTrust it

Can anyone check that it is true?

A source backs the walking claim.
04 · ContinuationWhat next?

Does it help with what they ask next?

Units, schools, price, availability.

These are things you control. Not ranking factors I am claiming.

Take one fact. How long it takes to walk to the station. Can anyone find it? Or is it locked inside a brochure, behind a form?

Does it make sense on its own? “Steps away” tells nobody anything. “Eight minutes, this route” tells them something.

Can anyone check it? Who measured it, and when. A number with nobody behind it is just a nicer mood.

Does it help with what comes next? After the walk they ask about the school, the unit, and the price.

A property-page example

Here are all four checks, on one page.

Hypothetical example
Promotional page

A life beyond extraordinary

Perfectly connected. Thoughtfully designed. Created for every chapter.

  • Facts buried
  • No dates
  • Nothing to compare
→
Evidence surface

Family-focused launch near rail: key facts

Access: public fact blockExtraction: complete rail statementSupport: source + checked dateWhat next: units, price + availability

*Illustrative only. Verify route, date, source, and current availability.

  • Answers the question
  • Dated, sourced facts
  • Points at what is next

Diagnose before rewriting

Fix the thing you actually found.

It is rightRail and price facts are current
→Do nothing. Watch it.
It is wrongOld price or old handover date
→Fix where it came from
It is buriedRail detail hidden in a brochure
→Move it into the page
It is missingNo checked walking route
→Add the one fact
Nothing answers itNo page speaks to the question
→Now you build

Doing nothing is also a real answer.

Facts are right? Then leave it alone. Watch it. Do not invent work to keep a system busy.

Price is old? Fix where it came from. Rewriting the page just moves the wrong number somewhere prettier.

Buried in a brochure? Move it into the page. You do not need more words. You need the words somewhere findable.

No checked route? Add that one fact. That is the whole job. One line.

Nothing answers the question at all? Now, and only now, you build something new.

Part 3 · Run the loop · Inside my own setup

An AI marketing team is a set of roles—not a room full of chatbots.

Photograph this
Plain software No judgement. Never tired.
  • Collects the data
  • Checks it is valid
  • Routes the work
  • Stops duplicates
  • Holds the state
  • Runs on schedule
  • Shouts when output goes stale
Agents, on a leash Judgement inside limits.
  • Reads the opportunity
  • Researches and verifies
  • Writes the brief
  • Drafts or restructures
  • Diagnoses the technical fault
  • Reviews somebody else’s work
  • Reports and recommends
A named person Accountable. No exceptions.
  • Every factual property claim
  • Anything published
  • Anything sent to a customer
  • Any change to customer data
  • Budget and spend
  • Strategic trade-offs
  • Anything with legal or brand risk

The reviewer is never the one who wrote it.

One end-to-end workflow

Same loop as before. Now look at who does what.

1 · SoftwareSpots itWe are not in the answer
→
2 · AgentWorks out whyRail fact is buried
→
3 · AgentDrafts the fixOne sourced line
→
4 · Another agentChecks itRoute and source
→
5 · A personSigns offName attached
→
6 · SoftwareAsks againSame question, later
Every week. Whether anyone remembers or not.
SoftwareAgent judgmentHuman authority

Software collects. Agents draft. People approve.

Software spotted it. Software does not get bored and it does not forget. That is the only thing it is better at than you.

An agent worked out why. The rail fact was there. It was just buried in a brochure.

An agent drafted the fix. One line, with a source.

A second agent checked the first. The one who wrote it does not get to mark its own work.

A person signed it off. Their name is on it. That is the only box on this slide with a name in it.

Then software asked again. And look where it ends. Same place it started.

Every rule came from a mistake

Every mistake showed us where a rule belonged.

01Source lineage

Every material claim needs a named source and an independent check.

02Durable operations

Every run needs stable handoffs, records, and stale-output alerts.

03Human authority

Every consequential action needs a named accountable owner.

It made a number upSounded right. No source anywhere.
→Named source, checked by someone else
Work went missingToo much started at once. Nobody owned it.
→Limit what runs. Hand over properly.
Nothing happened, quietlyLaptop slept. The run never ran. Nothing told me.
→Keep a record. Shout when it goes stale.

Safe does not mean the system does less. It means you can see it, stop it, and fix it.

An agent handed me a number that sounded completely right. There was no source anywhere. So now every claim needs a named source, and somebody other than the writer checks it.

We started too much work at once and ownership just dissolved. So now there is a limit on how much runs at a time, and a handover that cannot drop.

My laptop went to sleep. The run never happened. And nothing told me. So now the alarm goes off when something expected does not happen. Silent failures do not throw errors.

Your turn

Turn one buyer question into one workflow.

3 minutes
1One question

What would a real buyer ask?

“Which launch near rail fits a young family?”
2One evidence gap

What fact is missing or weak?

Walking time is not verified.
3One workflow

What recurring task could the system run?

Weekly answer audit.
4One human decision

What must remain accountable?

Approve every public property claim.

No worksheet? Use your notes app or a blank page.

Your Monday-morning move

Start with one question, one workflow, and 30 days.

Day 1Save the baseline answer
→
Week 1Fix one evidenced gap
→
Day 30Repeat the test and compare
Buyer questionWhich new launch near an MRT suits a family with young kids?→A shortlist they can trust

So. What do you test first?

Questions