TPTK Workflow Lab TWF-001 – controlled test – July 15, 2026
Both AI flyers kept the listing facts, but neither final followed every instruction
Direct answer: ChatGPT and Gemini Nano Banana Pro each preserved all 12 of 12 grouped copy requirements in this controlled AI real estate flyer test. Nano Banana Pro objectively followed the no-added-elements and page-proportion constraints more closely. TPTK’s editorial assessment is that the OpenAI flyer used the more polished and ambitious layout. Neither consumer correction fixed its first-pass issue categories, so this test does not produce an instruction-complete client flyer or a declared personal winner.
The two consumer finals, side by side
These are Si’s preserved final outputs, presented without added labels over the artwork. The captions carry the analysis so the evidence remains visually intact.
ChatGPT consumer final
More ambitious composition, two unresolved constraints
The layout polish is TPTK editorial judgment. Objectively, the output stayed 1024 x 1536 instead of 8.5:11 and used a stylized car in the EV-ready feature icon despite the no-vehicle instruction. That icon is an instruction deviation, not an invented textual property claim or a vehicle inserted into the property photo.
Gemini Nano Banana Pro final
Closer constraint fit, one typography caveat
The 800 x 1024 canvas was within 1.103% of the target proportion and no prohibited addition was observed. Several numeral 1 glyphs resembled a capital I in three grouped areas: the property-facts line, open-house details, and contact line. The intended values were understandable, but the raster image cannot prove the underlying character.
Evidence-led verdict
OpenAI produced the more polished and ambitious layout in TPTK’s editorial assessment; Nano Banana Pro followed the no-added-elements and page-proportion constraints more closely on the measured evidence. That is a tradeoff, not a personal-winner claim. Si did not provide a subjective preferred client flyer for this publication, and the article does not infer one.
Objective scorecard: accuracy, proportion, additions, legibility, and correction
ChatGPT final
- Required copy
- 12/12 groups; 0 property-fact errors; 0 feature errors
- Canvas
- 1024 x 1536; ratio 0.666667; -13.725% from target
- Additions
- One stylized car in an EV-ready icon
- Legibility
- 0 malformed groups; 0 glyph-ambiguous groups coded
- Correction
- 0 of 2 first-pass issue categories resolved
- Instruction-complete
- No
Nano Banana Pro final
- Required copy
- 12/12 groups; 0 property-fact errors; 0 feature errors
- Canvas
- 800 x 1024; ratio 0.781250; +1.103% from target
- Additions
- None observed
- Legibility
- 0 malformed groups; 3 glyph-ambiguous groups coded
- Correction
- 0 of 2 first-pass issue categories resolved
- Instruction-complete
- No
Target width-to-height ratio: 8.5 / 11 = 0.772727. Relative ratio error is calculated as (observed ratio – target ratio) / target ratio x 100.
The property photo remained geometrically consistent in every output
No material change to the house, landscaping, sky, view, or visible property condition was observed in the four consumer outputs or the two replication outputs. Feature matching and homography estimation produced inlier rates from 91.86% to 96.31%, with estimated source-image coverage from 97.79% to 99.16%.
How TWF-001 was run
- A conspicuously fictional property source image and 12 grouped copy requirements were frozen before testing.
- Si submitted the same first prompt to the OpenAI consumer app and Gemini Nano Banana Pro.
- The first output from each tool was preserved before the same correction prompt was submitted.
- Si preserved the two corrected consumer finals and the visible tool-surface screenshots.
- Codex audited copy, dimensions, additions, typography, correction behavior, and source-image continuity from the frozen evidence.
- A separate OpenAI replication used the frozen source and prompts. It is reported independently and does not replace Si’s consumer run.
Why 12 copy groups?
The audit grouped inseparable lines rather than counting every word as its own requirement: fictional-status line; address; price; headline; property-facts line; five individual feature lines; open-house block; contact line; and the fictional-use footer. All outputs preserved all 12 groups. A group was also checked for exact facts, omissions, substitutions, malformed text, and ambiguous glyphs.
Identical first prompt
Use a bright, modern Pacific Northwest editorial style with excellent typography, strong hierarchy, generous spacing, and restrained teal, coral, and deep navy accents. The flyer should feel credible enough for a professional real-estate agent to review, not like a generic template. Keep all text horizontal and highly legible. Do not use tiny text except for the required footer.
Include this exact copy with no omissions, substitutions, invented facts, or spelling changes:
FICTIONAL WORKFLOW TEST – NOT A REAL LISTING
1847 Harbor Pine Lane, Cedar Bay, WA
$725,000
Pacific Northwest comfort with room to work from home
3 bedrooms | 2.5 bathrooms | 2,140 sq ft | 0.24 acre | Built 2018 | 2-car garage
Water-view deck
Dedicated home office
Heat pump heating and cooling
Fenced backyard
EV-ready garage outlet
OPEN HOUSE
Saturday, July 18 | 11:00 AM-2:00 PM
Jordan Lee | Cedar Bay Realty | (555) 014-0188 | jordan@example.com
Fictional property and contact information created solely for AI workflow testing.
Return one complete flyer image. Prioritize factual accuracy and readable text over decoration.
Identical correction prompt
The correction prompt changed the designs, but not the coded issue categories
ChatGPT first pass
The first pass already had all 12 copy groups, but it used a 2:3 canvas and the prohibited car icon. Both categories remained in the consumer final.
Nano Banana Pro first pass
The first pass had all 12 copy groups and no prohibited addition. Its near-target ratio and numeral-1 ambiguity both remained in the consumer final.
Correction effectiveness here means whether the frozen, observable issue categories were resolved. It does not mean the two images were visually identical. Both tools redesigned parts of the flyer; neither reached an objective category-resolution rate above 0% in the consumer run.
Separate OpenAI replication: ratio improved sharply, car icon remained
Replication evidence, not Si’s consumer run
Codex completed a controlled OpenAI replication with the frozen source and the two frozen prompts. The first pass was again 1024 x 1536. The corrected output changed to 1103 x 1426, a ratio of 0.773492, which is within +0.099% of the 8.5:11 target. It was near-exact, not mathematically exact. The stylized car in the EV-ready icon remained, and copy stayed at 12/12.
The replication supports a narrower conclusion: the OpenAI workflow can respond materially to the page-proportion correction, but this replication still did not satisfy every instruction. It does not establish universal reliability and is not evidence of Si’s preference.
What a real estate professional should check before a flyer reaches a client
Neither final should be treated as client-ready on objective compliance alone. This test did not include broker, MLS, printer, seller, buyer, regulator, or client review. A useful production gate would include:
- Facts: verify every address, price, feature, measurement, date, time, and contact field against the approved source.
- Photo truthfulness: compare the output with the original property image at useful zoom and reject altered conditions or invented views.
- Page production: confirm exact dimensions, print margins, bleed, export resolution, and whether the printer changes scaling.
- Typography: inspect every numeral, separator, email address, and phone number rather than relying on overall readability.
- Unauthorized visuals: remove icons or decorations that violate the brief even when they do not create a textual claim.
- Rules and approvals: apply current brokerage, MLS, advertising, disclosure, accessibility, and client-approval requirements for the actual market and distribution channel.
Download the evidence and structured scorecard
The source, consumer outputs, replication outputs, and surface screenshots are direct media files. The JSON and CSV scorecards are generated in your browser from the embedded frozen data because this WordPress installation rejected native JSON, CSV, and plain-text media attachments. No public attachment URL or DataDownload schema is claimed.
File-integrity note: the scorecard SHA-256 values identify the preserved original evidence files. WordPress re-encoded the two JPEG Gemini delivery copies while retaining their 800 x 1024 dimensions, so their public media bytes do not share the originals’ hashes. The preserved originals remain in the TWF-001 evidence package.
ChatGPT first passPNG – primary consumer evidence
ChatGPT finalPNG – primary consumer evidence
Nano Banana Pro first passJPG – primary consumer evidence
Nano Banana Pro finalJPG – primary consumer evidence
OpenAI replication firstPNG – separately labeled evidence
OpenAI replication correctedPNG – separately labeled evidence
OpenAI surface screenshotPNG – visible consumer configuration evidence
Gemini surface screenshotPNG – Gemini Pro, Images, Thinking evidence
Limits of this test
- This is one controlled consumer run per tool, not a representative benchmark of every account, model route, prompt, property type, or future product version.
- The Gemini image-model name is Si-confirmed as Nano Banana Pro. The visible configuration was Gemini Pro with Images enabled and Thinking selected.
- The OpenAI consumer wording is bounded to the app default Instant setting and Si’s report that he did not change the selection. A hidden image-model backend is not established.
- No generation times were preserved, so this article makes no speed claim.
- TPTK’s layout-polish statement is editorial judgment. Copy counts, dimensions, additions, glyph observations, hashes, and geometric matching are objective audit fields.
- Si’s personal preferred client flyer and 1-to-10 subjective ratings were not provided and are not inferred.
Frequently asked questions
Can ChatGPT create an accurate real estate flyer?
In this one controlled run, ChatGPT preserved all 12 grouped copy requirements and the source composition. It still missed the requested 8.5:11 proportion and retained a prohibited car icon, so the result was not instruction-complete.
Did Gemini Nano Banana Pro follow the brief better?
It followed the no-added-elements and page-proportion constraints more closely in this run. All copy remained understandable, but the numeral 1 looked like a capital I in three grouped areas. That is a typography caveat, not a missing fact.
Did either tool alter the property photo?
No material property-condition change was observed. SIFT/homography matching found strong geometric continuity, but that does not prove pixel identity or rule out subtle tonal edits.
Did the correction prompt fix the problems?
Not in the two consumer runs: neither resolved its first-pass issue categories. In the separate OpenAI replication, the ratio improved to within 0.099% of the target while the prohibited car icon remained.
Can either flyer be sent directly to a client?
Not on objective compliance alone. Neither consumer final was instruction-complete, and this test did not include broker, MLS, printer, seller, buyer, or client review. A human should verify facts, contact details, branding, dimensions, disclosures, and property-photo truthfulness.
What should agents check before using an AI flyer?
Verify every listing fact against an approved source, compare the property image with the original, check page size and print margins, inspect every numeral and contact detail, remove unauthorized visuals, apply brokerage and MLS rules, and obtain the required human approvals before distribution.
How to cite or reuse this audit
Suggested citation:
Fisher, Si. "AI Real Estate Flyer Test: ChatGPT vs Gemini." TPTK Workflow Lab TWF-001, The Pro Tool Kit, July 15, 2026. https://theprotoolkit.com/ai-real-estate-flyer-chatgpt-vs-gemini-test/
TPTK’s original scoring method and coded scorecard are available under CC BY 4.0 with attribution. That license does not extend to tool outputs, product-interface screenshots, the synthetic property source image, trademarks, or third-party material. Preserve the one-run scope and replication boundary when citing the findings.