I tested secure website access with a live power-of-sale search. Here is the workflow, the mistake it made, and what REALTORS® must verify.
By Sushil Mishra, REALTOR®, Co-founder of Catch The Key and creator of The Real Tech
Published: August 27, 2026 | Last updated: August 27, 2026
ChatGPT Work can now continue tasks inside supported websites after you sign in securely. I tested it inside TRREB REALM, where it built an active power-of-sale search, corrected a filtering mistake, and returned 35 matches across Hamilton, Burlington, Brampton, and Mississauga. Useful? Yes. Hands-free? Absolutely not.

Can ChatGPT Work actually sign in to TRREB REALM?
ChatGPT Work can open a supported website and continue working after the user completes authentication. It does not independently know or enter the user’s credentials. In my test, ChatGPT opened REALM, paused for the TRREB sign-in, and resumed only after I privately entered my User ID, PIN, and verification code.
According to OpenAI’s browser documentation, credentials entered through its secure sign-in form go directly to the cloud browser. The model cannot see the username or password, and OpenAI says it does not store them.
That wording matters. Saying “ChatGPT logged into REALM for me” is catchy, but it is not completely accurate.
I authenticated. ChatGPT continued the work inside the authenticated session.
I did not type my User ID, PIN, password, or verification code into the conversation. No agent should.
OpenAI also says ChatGPT Work’s cloud browser is separate from the browser on your own computer. It does not automatically inherit your local tabs, cookies, saved passwords, extensions, or existing login sessions.
What did I ask ChatGPT to search in REALM?
I asked ChatGPT to find properties currently available for sale under power of sale. The final search included freehold and condominium properties, looked for “power of sale” in public remarks, excluded sold-conditional listings, and covered Hamilton, Burlington, Brampton, and Mississauga.
My first request was less precise:
Search the properties which are sold under power of sale.
ChatGPT interpreted that as a request for historical sold listings. That interpretation was reasonable because I had used the word “sold,” but it was not what I intended.
I corrected it:
I am looking for properties currently available for sale under power of sale.
That one sentence changed the task from researching completed sales to finding active opportunities.
It is a simple example of why agents must describe the intended listing status, not assume the system understands what they meant.
What happened during the live REALM search?
ChatGPT built the search in stages, responded to corrections, and checked the final filters. It also made one geography error while closing the municipality selector. The complete test produced four distinct result counts before arriving at the final 35 listings.
The Real Tech REALM Test Log
This is first-party test data from my supervised REALM session on August 26, 2026. Listing availability can change at any time.
Search stage | REALM count | What the count represented | What needed correction |
|---|---|---|---|
Initial phrase search | 220 | Freehold and condo records with “power of sale” in public remarks | Included some sold-conditional listings |
Status corrected | 203 | Available listings across the broader TRREB coverage area | Geography was still too broad |
Four-city search, first attempt | 36 | Hamilton, Burlington, Brampton and Mississauga, plus an accidental area | Toronto C06 had been added unintentionally |
Verified four-city search | 35 | Requested cities with the selected property and status filters | Final filters matched the instruction |
The 36-to-35 correction is the part agents should remember.
While ChatGPT was closing the municipality selector, Toronto C06 was accidentally added. ChatGPT later noticed the extra district in the search URL, removed it, and reran the search.
The final number was 35.
The tool made an error, detected it, and corrected it. A REALTOR® still had to understand what the correct search should contain.
Does the 35-listing result include every power of sale?
No. The 35 results were listings matching the filters used during that session, including the phrase “power of sale” in public remarks. That is not the same as proving that every lender-controlled or power-of-sale property in the four cities was captured.
Some listings may use different wording. Others may disclose relevant details only in brokerage remarks, schedules, attachments, or offer instructions. A listing might use phrases such as “as is, where is” without explicitly saying “power of sale.” That phrase can also appear in a context that does not mean the property is currently being sold by a lender.
The defensible statement is:
On August 26, 2026, REALM returned 35 available freehold and condominium listings in Hamilton, Burlington, Brampton, and Mississauga whose public remarks matched “power of sale,” after sold-conditional listings and an accidental Toronto district were removed.
That sentence is specific, dated, and limited to what the search actually established.
How reliable is ChatGPT for an MLS search?
ChatGPT Work can be useful for building and organizing an MLS search, but the output is only as reliable as the instructions, selected fields, live interface state, and final verification. It should be treated as a supervised browser operator, not as an authoritative replacement for REALM, TRREB data, or a REALTOR®’s professional judgment.
Here is what worked well:
It navigated the authenticated platform after I signed in.
It examined the available search controls.
It changed from historical sold records to currently available listings after clarification.
It included freehold and condominium properties.
It removed sold-conditional listings.
It narrowed a broad search to four municipalities.
It noticed and removed an unintended Toronto district.
It summarized visible results in a usable format.
Here is what still required a human:
Defining what “available” meant
Correcting the initial misunderstanding
Confirming which property classes belonged in the search
Checking that the status filters were complete
Verifying the four intended municipalities
Understanding that a public-remarks keyword search can miss relevant listings
Deciding whether any result was genuinely suitable for a client
ChatGPT can reduce interface work. It cannot assume professional accountability.
How should REALTORS® verify work completed inside REALM?
REALTORS® should verify the platform state before relying on any summary produced by ChatGPT. I use a five-part method called the REALM Verification Loop: Request, Examine, Audit, Limit, and Manually confirm. It turns a vague browser task into a reviewable professional workflow.
The REALM Verification Loop
Step | Agent action | Question to answer |
R: Request precisely | Define status, geography, property class, keyword field, exclusions, and output | Did I describe the search I actually want? |
E: Examine the live filters | Review every selected field in REALM | Do the visible filters match my instruction? |
A: Audit unexpected results | Investigate odd counts, unfamiliar areas, and mismatched property types | Is there an accidental filter or missing exclusion? |
L: Limit the claim | Describe only what the search proves | Am I overstating the completeness or meaning of the results? |
M: Manually confirm | Open relevant records and verify details before advising or sharing | Would I defend this information to a client or broker? |
The named framework is deliberately simple. If an agent cannot complete all five checks, the search is not ready to be relied upon.
What prompt should REALTORS® use for a REALM search?
A good REALM prompt should define the login boundary, listing status, search field, geography, property classes, exclusions, output format, and actions ChatGPT is not allowed to take. It should also require a final filter audit before ChatGPT summarizes the results.
Use this as a starting template:
Open TRREB REALM. I will complete the sign-in when prompted. Do not ask me to provide credentials in the chat.
Once I am signed in, search for residential properties currently available for sale in Hamilton, Burlington, Brampton, and Mississauga where the public remarks contain the exact phrase “power of sale.”
Include freehold and condominium listings. Exclude sold, leased, expired, suspended, terminated, and sold-conditional listings.
Before giving me the results, show me every active filter and verify that no other municipality or Toronto district was added accidentally.
Return the total count and a table showing address, municipality, list price, property type, bedrooms, bathrooms, current status, days on market, and the relevant wording from the public remarks.
Do not save the search, change any record, send information, download confidential material, or contact anyone without asking me first.
Adapt the prompt to the client and the purpose. Do not treat it as a substitute for reviewing the actual search screen.
Which authenticated tasks could REALTORS® test next?
The best early use cases are repetitive, reversible, easy to inspect, and unlikely to expose confidential information. Agents should begin with research and organization tasks, then expand only after they understand the tool’s behaviour and their brokerage’s policies.
Practical examples include:
Build a preliminary listing search
Describe the buyer’s basic criteria and have ChatGPT configure the first version of the search. Review every field before using or sharing the results.
Organize listing histories
Ask it to collect permitted dates, statuses, prices, and days-on-market information into a comparison table. Confirm that every record belongs to the correct property.
Prepare a comparable shortlist
Use it to organize potential comparables by distance, property type, size, sale date, and relevant features. The agent still chooses the true comparables and makes the adjustments.
Review CRM activity
Inside BoldTrail or another compatible CRM, it may be able to locate a contact, summarize an activity timeline, identify missing information, or prepare a follow-up draft. Do not let it send or change records without review.
Gather showing feedback
It may be able to collect feedback from an authenticated platform and group recurring themes for a seller update. Agents must distinguish direct feedback from ChatGPT’s interpretation.
Retrieve invoices and receipts
Advertising platforms, accounting tools, and brokerage portals often require repetitive navigation to retrieve documents. Start with non-confidential records and verify every downloaded file.
Prepare a website update
In a compatible content management system, it could add an approved draft, check links, or identify missing page information. Publishing should remain a deliberate human-approved action.
For more practical tools, visit The Real Tech tools directory. Brokerages looking for hands-on implementation can also review The Real Tech workshops and services.
What privacy and compliance risks should REALTORS® consider?
REALTORS® remain responsible for client confidentiality, privacy, MLS compliance, brokerage policy, and the accuracy of information they use. A secure login protects credentials from the model, but it does not make every authenticated workflow appropriate or eliminate the need for informed judgment.
Use these boundaries:
Never paste passwords, PINs, verification codes, or recovery codes into the chat.
Avoid identification, banking records, credit reports, employment documents, agreements, lockbox codes, and other sensitive client information unless the workflow has been properly reviewed and authorized.
Confirm TRREB rules, platform terms, brokerage policies, privacy obligations, and client consent requirements.
Require approval before sending, submitting, publishing, purchasing, modifying, or deleting anything.
Inspect the actual page state rather than trusting only the written summary.
Clear the cloud-browser session or sign out when continued access is unnecessary.
OpenAI says ChatGPT asks for confirmation before consequential actions. REALTORS® should still give explicit boundaries in the prompt and watch for any request that could affect a client, transaction, record, or public-facing page.
Who can currently use authenticated website access?
OpenAI currently says authenticated website navigation in ChatGPT Work is available on web and mobile for Plus and Pro users, subject to rollout and workspace settings. Its documentation says website sign-in is not available for Enterprise or Edu workspaces at this time.
Some websites can block automated browsers or require a CAPTCHA. A supported plan does not guarantee that every website or login method will work.
The feature was included in OpenAI’s August 25, 2026 changelog. Agents should check the current documentation because plan access, supported platforms, and rollout conditions can change.
Is ChatGPT Work inside REALM worth testing?
Yes, but start with one narrow, low-risk task. The practical value is not autonomous real estate. It is reducing repetitive navigation, filtering, copying, and organization while keeping the REALTOR® responsible for the criteria, interpretation, privacy, and final decision.
My test produced a usable result. It also misunderstood my first instruction and briefly selected the wrong geography.
That is the honest picture.
This tool is capable enough to save time and imperfect enough to require supervision.
The agents who benefit will not be the ones who hand over everything. They will be the ones who define the task precisely, inspect the work, and know when the machine’s answer does not match the market reality.
Frequently asked questions
Can ChatGPT see my TRREB password or PIN?
OpenAI says credentials entered through its secure sign-in form go directly to the cloud browser and are not visible to the model. Never type passwords, PINs, or verification codes into the conversation itself.
Does ChatGPT connect to my regular browser session?
No. ChatGPT Work uses a separate cloud browser. OpenAI says it does not automatically access your local tabs, cookies, browser history, saved passwords, extensions, or existing signed-in sessions.
Did ChatGPT find every power-of-sale listing?
No. It found 35 listings matching the selected status, geography, property-class, and public-remarks filters at that time. Listings using different wording or disclosures outside public remarks may not have appeared.
Can ChatGPT prepare a CMA in REALM?
It may help organize candidate comparables and property data. Selecting appropriate comparables, making adjustments, interpreting market conditions, and forming a pricing opinion still require a qualified REALTOR®’s judgment.
Should ChatGPT be allowed to change listings or client records?
Not without explicit review and authorization. Begin with read-only research and organization. Require confirmation before any action that sends, submits, publishes, edits, downloads confidential material, or changes a record.
Subscribe to The Real Tech
The Real Tech tests new technology from a working REALTOR®’s perspective. That means showing what worked, what failed, and what still needs a human watching closely.
Subscribe at blog.therealtech.ca for practical workflows, real tests, and direct explanations without the usual technology theatre.
Sushil Mishra is a REALTOR® with REMAX West Realty Inc., co-founder of Catch The Key, and creator of The Real Tech. He tests real estate technology in live working environments and translates it into practical systems for agents and brokerages.
Sources and disclosure
The REALM example is based on Sushil Mishra’s supervised test conducted on August 26, 2026. Search counts and listing availability can change at any time. This article is educational and does not replace TRREB rules, brokerage policy, legal advice, privacy guidance, or independent verification of MLS information.
