ChatGPT Home Valuation Accuracy: A 198-Home Study for Real Estate Agents

Can ChatGPT accurately estimate a home value? In Touchpoint Valuation’s September 2026 sample of 198 recently sold single-family homes, ChatGPT estimates were within 10% of the reported sale price for 108 homes (54.5%). Median absolute error was 8.9%. These are results for this sample and prompt, not a national accuracy rating for real estate AI.

For agents searching for ChatGPT for real estate, AI for real estate agents, or AI home valuation accuracy, the practical question is not whether a chatbot can produce a polished answer. It is whether the comparable sales and final home value estimate hold up when checked against a transaction.

Original research by Touchpoint Valuation. Published October 18, 2026. Sale dates: September 2 to 20, 2026. Estimates collected in September 2026. We sell automated home value email software, which creates a commercial interest readers should consider. This study did not test Touchpoint’s estimates, other chatbots, or dedicated automated valuation models (AVMs).

Jump to results · Read the exact ChatGPT prompt · See limitations · Full dataset · How to cite

Real estate AI vs. AVMs vs. a professional CMA

ChatGPT is a general-purpose AI chatbot; it is not the same thing as a dedicated real estate AVM. An automated valuation model estimates property value from structured property and market data. Some AVMs use machine learning, but an AVM does not have to be a conversational AI tool. This experiment tested ChatGPT’s response to a request for comparable sales, not AVM accuracy as a category.

A comparative market analysis (CMA) is an agent’s assessment of recent comparable transactions and the subject home. Asking ChatGPT to imitate that process does not make its output a professional CMA or appraisal. We did not run a head-to-head comparison with an agent CMA, appraisal, or commercial AVM, so this study cannot rank those methods.

For background, read what an AVM means in real estate and how to read a CMA.

ChatGPT home valuation accuracy: key findings

• ChatGPT's estimate was within 10% of the reported sale price for 108 of 198 homes (54.5%). The median absolute error was 8.9%.
• Accuracy was highest for homes bought from 2016 to 2020, where 18 of 25 estimates (72%) were within 10%. It was lowest for homes owned 20 years or more, where 9 of 25 (36%) were within 10%.
• In the three designated recent-purchase cohorts, estimates were too low in 48 of 75 cases. In the three designated longer-held cohorts outside the 2006 to 2008 group, estimates were too high in 50 of 75 cases. The 2006 to 2008 group was too low in 15 of 23 cases. These pooled cohorts exclude the separately assigned construction-year group.
• Homes bought near the 2006 market peak had a within-10% rate of 43.5%, the second lowest of any group.
• Homes built in 2000 or later were within 10% in 69.1% of cases, compared with 49.0% for homes built before 2000. Newer homes in the sample were also owned for shorter periods, so this study cannot separate the two effects.
• Nine estimates (4.5%) missed by 50% or more, and eight of those were too high. With those nine removed, the within-10% rate rises to 57.1%.

Study at a glance

1. Summary

The data. This study asked ChatGPT to estimate the market value of 198 single-family homes that sold between September 2 and September 20, 2026, and compared each estimate with the reported sale price. ChatGPT's estimate was within 10% of the sale price for 54.5% of homes and within 5% for 31.8%. The median absolute error was 8.9% and the mean absolute error was 15.0%. Observed accuracy varied by purchase-period group, ranging from 72% within 10% for purchases made from 2016 to 2020 down to 36% for homes owned 20 years or more. The direction of the error also depended on ownership length, with shorter-held homes tending to be undervalued and longer-held homes tending to be overvalued. Homes bought near the 2006 peak were the exception and tended to be undervalued.

In plain terms. When ChatGPT was asked what a house was worth, its answer was close about half the time. For a typical home in the study, the miss was around 9%, which would be roughly $39,000 on a hypothetical home priced at the sample’s $439,000 median. That illustration is not the median dollar error; the actual median absolute dollar error across the 198 homes was $35,750. ChatGPT did noticeably worse on homes the owner had held for decades. A polished explanation is not a measured confidence score. This study did not formally score the tone or confidence of the responses.

2. Why we ran this study

The data. Touchpoint Valuation makes software that sends automated home value emails on behalf of real estate agents. That gives us an interest in how homeowners understand value estimates, and readers should weigh our findings with that in mind. This study does not measure how many homeowners use AI chatbots for home values. It measures something narrower and testable. When a general-purpose AI chatbot receives the basic facts a real estate professional would start from, how close does its estimate come to the price the home actually sold for?

In plain terms. Asking a chatbot what your home is worth costs nothing and takes seconds, so it is a natural thing to try. The answer comes back in polished, complete sentences that sound authoritative. We wanted to know whether that tone is backed by accuracy. A reported sale price provides an observable transaction benchmark, although it is not a perfect measure of market value and may reflect unusual sale conditions.

3. Key terms

The data. The following definitions apply throughout this article.

• Sale price: the reported price for the home's sale between September 2 and September 20, 2026. This is the benchmark for every comparison.
• Signed error: ChatGPT's estimate minus the sale price, divided by the sale price, expressed as a percentage. A positive signed error means the estimate was too high and a negative signed error means it was too low.
• Absolute error: the signed error with the sign removed. It measures the size of the miss regardless of direction.
• Within 10%: an estimate whose absolute error is 10% or less. The same logic applies to "within 5%" and "within 20%."
• Median error: the middle value when all absolute errors are sorted from smallest to largest. Half of the estimates missed by less than the median and half missed by more.
• Mean error: the average absolute error. A small number of very large misses can raise the mean well above the median.
• Disclosure state: a state where home sale prices are part of the public record. In non-disclosure states, such as Texas, sale prices are generally not published, so they could not be used to check estimates.
• Comparative market analysis (CMA): the standard method real estate agents use to price a home, based on recent sales of similar nearby homes, often called "comps."

In plain terms. If a home sold for $400,000 and ChatGPT said $440,000, the estimate was 10% too high, which still counts as "within 10%." If ChatGPT said $360,000, it was 10% too low, which also counts. Throughout this article we lean on the median more than the average, because the median describes the typical case. The average gets pulled upward by a handful of answers that were very far off.

4. Study design and sample selection

The data. The researcher-supplied sample contained 198 single-family homes in 37 disclosure states. Homes in non-disclosure states were not included. The study began with 200 homes, and two were removed because their sale records could not be re-verified. Both came from the group of homes bought from 2006 to 2008, which therefore contains 23 homes, while every other group contains 25. Each home belongs to exactly one group. Five groups were defined by how long the current owner had held the home at the time of sale. Two separately assigned groups covered purchases made during unusual market conditions: the 2021 to 2022 price surge and the 2006 to 2008 market peak. The supplied group labels are retained for recalculation, not treated as exhaustive duration bands. Exact boundary dates and assignment precedence were not supplied. The January 2006 boundary potentially overlaps the long-held and 2006 to 2008 groups: row 170 has purchase year 2006 but is assigned to the long-held group. Its exact purchase date is needed to resolve that boundary; no row has been reassigned based on rounded years alone. The eighth group was selected by construction year (built before 2000) regardless of purchase date. Each row appears once, but cohort eligibility overlaps; these groups are not a clean partition by ownership duration. The "Built before 2000" cohort contains 25 separately assigned homes, while the all-sample construction comparison contains all 143 pre-2000 homes.

Durations are approximate and rounded to one decimal place. Exact purchase and sale dates are not in the published dataset; adjacent cohorts can share displayed boundary values. The purchase-date ranges above were supplied by the researcher and cannot be verified from purchase year alone.

In plain terms. We selected houses that had recently sold and noted what each one sold for. We then asked ChatGPT to value each house without telling it the price. Most of the groups are based on how long the owner had held the property, because valuing a home bought last year is a very different problem from valuing one bought in 1985. Two groups focus on people who bought during famously unusual markets, just before the 2008 crash and during the pandemic price surge. One group looks only at older houses. We left out states that keep sale prices private, because we did not have a consistently public sale-price benchmark for this sample, and we plan to cover those states in a separate study.

5. The prompt and the reasoning behind it

The data. The same prompt template was used for every home, with that home's details filled in:

I'm trying to estimate the current market value of a home. Here are the details:

Address: {address}

Square footage: {sqft}

Bedrooms: {bedrooms}

Year built: {year_built}

Do not look up, reference, or use this property's own sale price, listing price, or sale history, even if it appears anywhere in your search results. Instead, search for 3-5 comparable properties that have sold recently in the immediate area, similar in size, age, and bedroom count. Use those comparable sales to calculate an estimated value for this property, the way a real estate agent would prepare a comparative market analysis.

List the comparable properties you used (address, sold price, sold date), then end your response with a single final line in exactly this format, with no other text after it:

ESTIMATE: $<number>

Each element of the prompt was chosen for a specific reason. The four property facts (address, square footage, bedrooms and year built) match what an agent typically has on hand before pricing a home. The instruction to ignore the home's own sale and listing history exists because a chatbot with web access could otherwise find the reported sale price and repeat it, which would measure its ability to search and say nothing about its ability to estimate. The request for 3 to 5 comparable sales mirrors the standard CMA method, so the request resembles one part of an agent’s workflow, without testing a complete professional CMA. The fixed final line ("ESTIMATE: $") made it possible to extract every estimate in the same way without interpreting the text of the response. An earlier version of our plan used the address alone. We replaced it with the prompt above, because an address alone is not what a professional would start from. All estimates in this study come from the prompt shown above.

In plain terms. We were not interested in whether ChatGPT can find a sale price online. Sometimes it can, and that would tell us nothing. We wanted to see whether it can work out a price the way an agent does, by looking at similar homes that sold nearby. So we gave it the basic facts about each house and asked it to show which comparable homes it relied on, while telling it not to use the home's own sale record. The required last line simply let us read the final number from every answer in a consistent way.

6. Limitations of the prompt

The data. The prompt has limits that affect how the results should be interpreted. It omits lot size, condition, renovations, finishes, views and waterfront status, all of which an agent would normally consider, so some errors on unusual homes may reflect missing information as much as weak reasoning. It asks for a single number, while a professional CMA usually presents a price range. It instructs ChatGPT not to use the home's own sale history, but the responses cannot prove that the instruction was followed in every case. It asks for "current" market value, while the benchmark is a sale price from a specific closing date, and exact query timestamps were not supplied, so the timing gap cannot be audited. It uses one fixed wording, and different wording could produce different estimates. The ChatGPT model version and settings are not recorded in the dataset. Responses were collected through the ChatGPT chat interface, and the study did not control for the possibility that earlier questions in a conversation influenced later answers. Finally, 27 of the 198 homes were skipped in the first pass and requested again afterward. The supplied dataset retains one estimate per home but does not document whether retries followed missing answers or prior successful estimates, or how any response was selected. Because the homes had already sold when estimates were collected, this is a retrospective comparison, not a demonstrated test of presale forecasting.

In plain terms. The prompt gave ChatGPT the same starting facts an agent has on paper, but none of what an agent learns by walking through a home, such as whether the kitchen was remodeled or whether the yard backs onto a busy road. Homes that are unusual in some way were always going to be harder to value from the basics. These results describe how ChatGPT performs when given roughly what a homeowner would type in. They are not a measure of the best answer ChatGPT could produce with more detail.

7. Overall results

The data. Across all 198 homes, ChatGPT's estimate was within 10% of the sale price for 108 homes (54.5%). Under a tighter 5% threshold the count was 63 homes (31.8%), and under a wider 20% threshold it was 155 homes (78.3%). The median absolute error was 8.9% and the mean absolute error was 15.0%. Of the 198 estimates, 87 were above the sale price, 103 were below and 8 were equal to it. The median signed error was -1.3%, meaning the typical estimate was slightly low. The mean signed error was +1.4%, because the largest misses were mostly overestimates and they pulled the average upward. In the whole sample, high and low errors largely offset each other, but Section 9 shows that this balance hides a consistent pattern by ownership length.

In plain terms. About half of ChatGPT's answers were close to the real price, and roughly four out of five were within 20%. A small number were badly wrong. The typical answer missed by about 9%. The average miss was larger, around 15%, because a few very poor answers raised the average. When the median and the average are this far apart, the median is the better guide to what a homeowner should expect from a single answer.

8. ChatGPT real estate valuation results by purchase-period cohort

The data. The share of estimates within 10% of the sale price was highest for homes bought from 2016 to 2020 (72%) and from 2020 to 2023 (68%). It was lowest for homes owned 20 years or more (36%) and for homes bought near the 2006 peak (43.5%). Median error followed a similar pattern, at 5.8% and 5.3% for the two strongest groups compared with 14.8% and 11.2% for the two weakest. Homes bought from 2023 to 2026 had the lowest mean error (8.7%) and the highest within-20% rate (96%), but only 56% within 10%. That combination indicates that estimates for recent purchases were rarely far off but often moderately off.

In plain terms. ChatGPT did best on homes the owner had bought roughly 3 to 10 years before selling, where about seven in ten answers were close. It did worst on homes owned for 20 years or more, where only about one in three answers was close and nearly half missed by more than 20%. One possible explanation, which this study does not test, is that a home owned for decades has had more time to change, through renovations or years of wear, and is more likely to differ from the nearby homes ChatGPT compares it with. Recent purchases showed a different pattern: ChatGPT was rarely far off, but it often missed by a moderate amount.

9. Direction of the errors

The data. For the three designated recent-purchase cohorts (75 homes in total), ChatGPT's estimate was below the sale price in 48 cases and above it in 23, with a median signed error of -3.6%. For the three designated longer-held cohorts, excluding both the 2006 to 2008 group and the separate construction-year group (75 homes in total), the estimate was above the sale price in 50 cases and below it in 23, with a median signed error of +5.4%. Homes bought near the 2006 peak did not follow this pattern. Despite being owned 18 to 21 years, they were estimated too low in 15 of 23 cases, with a median signed error of -7.5%. The built-before-2000 group, which mixes recent and long-term owners, was estimated too low in 17 of 25 cases.

In plain terms. ChatGPT tended to put recently bought homes below what they actually sold for, and to put long-owned homes above it. Because those two mistakes run in opposite directions, the overall average looks balanced even though individual answers often were not. The exception was people who bought just before the 2008 crash, whose homes ChatGPT usually priced too low. For an agent, this means these sample-level differences are worth discussing, but they cannot predict whether an individual seller’s AI estimate is high or low.

10. AI property valuation for homes bought near the 2006 peak

The 23 homes in the January 2006 to December 2008 purchase cohort had a within-10% rate of 43.5% and a median absolute error of 11.2%. Estimates were below the sale price in 15 of 23 cases, with a median signed error of -7.5%.

This group did not follow the tendency toward overestimation seen in the other longer-held cohorts. The study does not establish why. Purchase timing is a grouping variable here, not proof of the mechanism behind an AI valuation error.

11. AI home value estimates for 2021 to 2022 purchases

The 25 homes purchased from January 2021 to November 2022 had a within-10% rate of 52% and median absolute error of 9.3%. Fourteen estimates were too low and 11 too high. The median signed error was -1.7%.

The mean signed error of +5.1% is strongly affected by one estimate that was 126.2% too high. Without that home, the group’s mean signed error was approximately +0.1%. Local sale conditions, property details, and the selected comparables need examination before attributing this result to a particular market cycle.

Market context, not an explanation of the errors: S&P Dow Jones Indices reported that the national home price index rose 18.8% during calendar 2021. Its July 2026 report showed +1.93% national year-over-year growth, with Chicago at +6.86% and Seattle at -1.57%. Those are different time comparisons; they do not establish cumulative inflation-adjusted appreciation since the 2021 to 2022 purchases or explain this sample’s errors.

12. Construction year

The data. Across all 198 homes, the 55 homes built in 2000 or later had a within-10% rate of 69.1% and a median absolute error of 6.5%. The 143 homes built before 2000 had a within-10% rate of 49.0% and a median absolute error of 10.1%. The dedicated built-before-2000 group (52% within 10%, median error 9.5%) performed close to the overall sample average, but this comparison understates the difference, because 72.2% of the full sample was also built before 2000. Construction year and ownership length are linked in this sample: a home built after 2000 cannot have been owned for more than about 26 years, and many newer homes in the sample were bought recently. This study cannot determine how much of the gap comes from the age of the building and how much comes from the length of ownership.

In plain terms. ChatGPT did noticeably better on newer homes. About seven in ten homes built in 2000 or later got a close estimate, compared with about half of older homes. The catch is that newer homes have usually changed hands more recently, and recent purchases were easier for ChatGPT in general. So it is fair to say newer homes were easier to value in this study, but we cannot say whether that was because of the house itself or because of when it was last bought.

13. The outliers

The data. Nine of the 198 estimates (4.5%) missed the sale price by 50% or more. Eight were too high, and seven of those eight were homes that sold for less than $400,000. One was too low, a home that sold for $1,000,000. Eight of the nine homes were built before 1990. With these nine estimates excluded, the within-10% rate rises from 54.5% to 57.1% (108 of 189) and the mean absolute error falls from 15.0% to 11.7%. The median absolute error is less affected by outliers, which is one reason we use it as the main measure of a typical miss.

In plain terms. Most of the very large misses were lower-priced homes that ChatGPT valued far too high. The one large miss in the other direction was an expensive home that ChatGPT valued at less than half its sale price. This study cannot say why these particular homes were missed so badly. Some of these sales may have involved circumstances that public records do not show, such as a sale between relatives or a home needing major repairs, and no estimate built from basic facts could anticipate those. These are the estimates reported in the supplied study dataset. A discrepancy that large deserves investigation before anyone uses the number to set a list price.

14. How real estate agents can respond to a seller’s ChatGPT estimate

The data. A homeowner who has asked a chatbot for a value arrives at a listing conversation already holding a number. In this study, that number was within 10% of the reported sale price 54.5% of the time, and less often for long-term owners (36%) and for owners who bought near the 2006 peak (43.5%). The direction of the error is also informative. In the three longer-held ownership groups outside the 2006 to 2008 cohort, estimates were more often above the reported sale price, which can make a realistic list price harder to accept. Recent buyers, and owners who bought from 2006 to 2008, were more likely to receive an estimate below it, which can make them hesitant to list at all.

In plain terms. When a seller mentions a ChatGPT number, it works better to ask how they got it than to dismiss it. From there, the agent can walk through the homes that actually sold nearby and explain how the seller's home compares, including the details a chatbot cannot see, such as condition and updates. Treating the AI figure as one starting point that deserves checking respects the seller's research and positions the agent as the person who can explain it.

The data. The practical effect on the listing presentation is that the pricing discussion has often started before the agent arrives. A seller whose chatbot figure is too high may resist a realistic list price, and a seller whose figure is too low may delay a sale that would make sense for them. In our view, both situations are easier to handle when the agent has already been in regular contact with the homeowner and has been sending credible, local market information over time. This is our opinion and was not tested in this study.

In plain terms. Homeowners will keep looking up numbers online. Clear, local updates can give an agent a reason to discuss a surprising estimate. This study did not measure replies, trust, leads, or conversion.

15. Limitations of the study

The data. The sample of 198 homes is modest, and most groups contain 25 homes, so a single group's percentage could move by several points with a different selection of properties. Each home was estimated once. Chatbots can give different answers to the same question, and this study did not measure how consistent ChatGPT's estimates are from one run to the next. The sample includes only single-family homes in disclosure states. Only ChatGPT was tested. Sale prices reflect what a particular buyer and seller agreed on, which can include unusual transactions. The 10% threshold was chosen for this study and is not presented as an industry standard, and other thresholds produce different percentages, as shown in Section 8. The limits of the prompt itself are described in Section 6.

In plain terms. This is a careful look at 198 real sales, not a final verdict on AI home valuations. In this sample, ChatGPT was within the selected 10% threshold about half the time and had lower accuracy in the longest-held group. These are descriptive findings, not a statistically established national pattern. The exact percentages could shift in a larger study. We plan to repeat this research with more homes and multiple runs per home.

How to use AI for real estate pricing conversations responsibly

A suggested seller script, not a tested study outcome: “Thanks for sharing that estimate. Can we look at the comparable homes it used together? I’ll check the sale dates, location, condition, and how those homes compare with yours before we decide on a list price.”

• Ask for the original ChatGPT prompt, response, and date rather than only the final number.
• Verify each comp’s address, sold status, closing date, and reported price against sources you can legitimately access.
• Check size, lot, bathrooms, renovations, property condition, and location differences the prompt did not include.
• Prepare your own CMA. Do not reuse an unverified AI estimate as an appraisal or promise of a sale price.
• Separate a home value estimate from home equity. Equity requires the current loan balance; net proceeds also require selling costs and other obligations.

If you send recurring home value emails to past clients, label the estimate and its limitations clearly. Touchpoint’s approach starts with the purchase price and date entered by the agent, then adjusts for local home price changes. It is a market-data estimate, not an appraisal, broker price opinion, or CMA. Nothing in this study proves that Touchpoint is more accurate than ChatGPT.

16. ChatGPT, AI real estate valuation, and AVM questions

How accurate is ChatGPT at estimating home values?

In this study of 198 recently sold single-family homes, ChatGPT's estimate was within 10% of the sale price 54.5% of the time. The typical (median) miss was 8.9%.

Is ChatGPT more accurate for some homes than others?

Yes. It was most accurate for homes bought from 2016 to 2020 (72% within 10%) and least accurate for homes owned 20 years or more (36% within 10%). Homes built in 2000 or later were also estimated more accurately than older homes.

Does ChatGPT overvalue or undervalue homes?

Both, depending on the home. In the shorter-held ownership cohorts, estimates were more often too low; in the three longer-held cohorts outside 2006 to 2008, they were more often too high. This is a sample-level pattern, not a prediction for every home. Homes bought from 2006 to 2008 were an exception and were usually estimated too low.

What information was ChatGPT given?

Each home's address, square footage, number of bedrooms and year built. ChatGPT was told not to use the home's own sale or listing history and to base its estimate on 3 to 5 comparable nearby sales.

How was accuracy measured?

Each estimate was compared with the home's reported sale price. The error is the difference between the estimate and the sale price, expressed as a percentage of the sale price.

Should homeowners rely on ChatGPT for a home value?

Based on this study, a ChatGPT estimate is a rough starting point at best. It was close about half the time, and a local real estate professional who can see the home's condition and nearby sales is better placed to give a reliable figure.

Is ChatGPT an automated valuation model (AVM)?

Not in the sense tested here. This study used a general-purpose chatbot prompted to find comparable sales. It did not test a dedicated AVM, and its results should not be used as an accuracy score for all automated valuation models.

Can ChatGPT replace a real estate agent’s CMA?

This study does not establish that. It did not compare ChatGPT with an agent CMA, and the prompt omitted condition, renovations, lot size, and other property-specific information. Verify the comparable sales and prepare an independent pricing analysis.

Does this study show which AI is best for real estate?

No. Only ChatGPT was tested, the model version was not recorded, and each retained home has one reported estimate. There is no comparison across AI tools or AVM providers.

17. How to cite this study

The data. Suggested citation: Touchpoint Valuation. (2026, October). ChatGPT Home Valuation Accuracy: A 198-Home Study for Real Estate Agents. The full property-level dataset appears in the appendix below, with homes identified by city and state only.

In plain terms. You are welcome to quote or reference these findings. Please link back to this page so readers can see the full method and data.

18. Data source and reproducibility

Primary-source market context: S&P’s calendar 2021 home price release, February 22, 2022; S&P’s historical 2006 to 2012 market summary, February 26, 2013; S&P’s July 2026 home price release, September 29, 2026. These sources provide market context, not validation of individual transactions or ChatGPT responses.

Source: Touchpoint Valuation’s supplied ChatGPT home valuation accuracy dataset, collected September 2026. The full property-level data appears in the appendix below.

The appendix supports recalculating errors from the reported sale prices and estimates. It does not independently verify the underlying transactions. Full addresses, exact transaction dates, original ChatGPT transcripts, verified comparable-sale records, model version, and settings are not published here. Because of those omissions, readers cannot fully reproduce the original estimates or audit whether ChatGPT avoided the subject property’s own sale price. The sampling procedure is not documented in sufficient detail to establish random selection or national representativeness; the groups were deliberately balanced rather than weighted to the US housing market.

The sale prices and property descriptions are researcher-supplied records, not independently authenticated by this publication workflow. Reported results were recalculated from the supplied rows using unrounded prices and estimates. Displayed percentages are rounded. There was no formal significance testing or control for confounding variables. Small cohort differences should be treated as exploratory.

Appendix: Full dataset

The data. The table lists all 198 homes in the study, identified by city and state only. Years owned is measured from the purchase date to the sale date. Signed error is ChatGPT's estimate minus the sale price, divided by the sale price, so a positive value means the estimate was too high.

In plain terms. Every supplied row is listed here so readers can recalculate the figures. City and state alone are insufficient to verify the underlying sale records. A plus sign means ChatGPT guessed too high and a minus sign means it guessed too low.