Is AI Actually Good at Real Estate Comps Yet?

CompsAgent Team · August 21, 2026 · 12 min read

Start with what agents are actually doing

This isn't a hypothetical debate any more. A February 2026 RPR/NAR survey found 82% of agents use AI daily, with 34% reporting savings of four or more hours a week.

But look at what those hours are going into — listing descriptions, follow-up emails, market summaries, CMA narratives. Writing tasks. The adoption is real and the value is real, and almost none of it is valuation.

The gap between "82% use it daily" and "it can price a house" is where agents get into trouble.

The failure mode nobody warns you about

Ask a general-purpose chatbot for "recent comparable sales near 1234 Oak Street" and you will usually get an answer. It will have addresses. It will have sale prices and dates. It will be formatted like a comp table.

A meaningful share of it may be fabricated.

This isn't the model malfunctioning — it's the model doing what it was built to do. A language model generates the most plausible continuation of text. Asked for comps with no data access, the most plausible continuation is a list of things that look like comps. Street names in that area, prices in that range, dates in that window. It has no mechanism to distinguish "this sale happened" from "this sale is the kind of thing that happens here."

An AVM that's wrong gives you a bad number. A chatbot that's wrong gives you a bad number attached to a street address that doesn't exist — which is a considerably worse thing to hand a seller who knows the neighborhood.

The specific danger is that fabricated comps are more convincing than a bad AVM estimate, because they arrive with the apparatus of evidence: an address, a date, a price. A seller can't tell the difference. Neither can you, unless you check every one against the MLS — at which point you've done the work anyway.

Two different technologies, constantly confused

Most confusion here comes from "AI" covering two unrelated things:

	AVMs (statistical models)	LLMs (chatbots)
Examples	Zestimate, Redfin Estimate, lender AVMs	ChatGPT, Claude, Gemini
Built to	Predict a sale price from property data	Generate and reason over text
Data access	Direct, structured, licensed	None by default
Good at	Typical homes in data-rich markets	Explaining, summarising, structuring
Bad at	Unusual property, thin markets, condition	Producing facts it wasn't given
Failure looks like	A number that's off	Confident, well-formatted fiction

AVMs are genuinely good at what they do within their limits. Zillow's own published accuracy — around 1.74% median error on-market and 7.20% off-market — is a real achievement, and that off-market figure is also exactly why it can't price a listing on its own.

LLMs aren't worse valuation engines than AVMs. They aren't valuation engines at all. Judging them on price accuracy is like judging a word processor on arithmetic.

Where AI genuinely earns its place

Four uses that hold up, all with the same shape — language work around verified data:

1. Cross-checking comps you already have

Give a model comps from a real source and have it search public listings for the same properties. Now it's verifying rather than inventing, and a discrepancy between the API price and the public listing is a genuinely useful flag.

This is what our own pipeline does — the comps come from real data providers, and a separate model pass web-searches public listings to sanity-check the prices. Where a real second source disagrees materially, the report shows the reconciliation and marks the comp rather than silently swapping a number. The distinction we hold to internally is worth stating: blending multiple sourced, cited numbers is a different thing from letting a model produce one.

2. Writing the narrative around settled numbers

Once comps are selected and adjustments computed, someone has to write two paragraphs explaining what the data means. That's a language task, and models are good at it.

The safeguard is strict separation: the narrative describes numbers that were already decided; it never contributes one. In our own build that's a hard architectural line — the component that writes prose can't influence a price.

3. Summarising market context

Turning absorption rate, DOM and sale-to-list into a readable paragraph for a specific price band. Again — you supply the figures, it does the prose.

4. Explaining an adjustment to a client

"Why is this comp adjusted down $18,000?" is a communication problem. Models are good at plain-language explanation, and you can check the output because you already know the answer.

Where it still fails

• Sourcing comps. Covered above. The single highest-risk use.
• Selecting comps. Even with real data, a model doesn't know the school boundary runs down that street or that the "comparable" backs an arterial. Screening is local judgment.
• Condition. It hasn't seen the kitchen. Nothing fixes that except the walkthrough.
• Micro-markets. Two blocks can behave differently and no general model knows that.
• Non-disclosure states. In Texas and similar states, sale prices aren't in public records at all — so a model relying on public data has less to work with than agents assume.
• Being accountable. You sign the CMA. A model's confidence is not a defense.

If you're going to use a chatbot anyway

Realistically, many agents will. Rules that keep it safe:

1. Never ask it to find comps. Paste in comps from your MLS and ask it to reason about them.
2. Ask for the adjustment logic, not the number. "What should I be adjusting for between these two?" is a good question. "What's this worth?" isn't.
3. Verify every factual claim. If it names an address, price or date you didn't supply, check it before it leaves your desk.
4. Prefer tools with real data access over a general chatbot. A model connected to an actual comp source is a different proposition from one guessing.
5. Never paste client personal or financial information into a general consumer tool.

The honest state of play

AI has meaningfully changed CMA production — just not the part people expected. It's compressed retrieval, formatting, layout and narrative writing, which is most of the three and a half hours a manual CMA takes.

It has changed comp selection and condition assessment approximately not at all. Those still need someone who knows the market and has seen the house.

Which means the job hasn't been automated — it's been re-weighted. Less of an agent's time on assembly, the same amount on judgment, and judgment is now a larger share of what the agent is contributing. That's a better job, not a smaller one.

The bottom line

Is AI good at real estate comps yet? It's good at everything around comps — explaining, checking, summarising, formatting — and bad at the part where the comps come from. Used for reasoning over data you supply, it's a genuine time saver. Used as a data source, it produces confident fiction that's more dangerous than an obviously wrong estimate, because it looks like evidence.

The agents getting real value from it are the ones who understood that distinction early and built their process around it.

Frequently asked questions

Can ChatGPT find real estate comps?

Not reliably. A general chatbot has no MLS access and no comp database, so when asked for comparable sales it will often produce addresses and prices that look plausible and are not real. It can reason about comps you supply; it cannot source them.

How many real estate agents use AI?

A February 2026 RPR and NAR survey found 82% of agents use AI daily, with 34% reporting savings of four or more hours per week. The heaviest uses are listing descriptions, follow-up emails, market summaries and CMA narrative writing rather than valuation itself.

Is AI good at property valuation?

Statistical models — AVMs like the Zestimate — are genuinely good at valuation when a property is typical and data is plentiful, and unreliable on unusual properties or in thin markets. Language models are a different technology and are not valuation engines at all; their value is in explanation and structure, not in producing a number.

What is AI genuinely useful for in a CMA?

Cross-checking comps you already have against public listings, writing the narrative around numbers that were already computed, summarising market context, and drafting the explanation of an adjustment. All of these are language tasks around verified data rather than sources of the data itself.

Will AI replace the CMA?

It is already replacing the assembly work — retrieval, formatting, layout and narrative drafting. It is not replacing comp selection or condition assessment, which depend on local knowledge and on having walked the property.

Related reading

• How Accurate Is a Zestimate? The 2026 Numbers
• How to Find Real Estate Comps Without Guessing
• Are Automated CMAs Accurate?
• MLS Data vs. Public Listing Aggregators