Why Might Your Radar Score Move This Week?
On 16 October 2026, Radar moved its ChatGPT measurements from gpt-4o-mini to gpt-6-luna, because OpenAI is retiring the gpt-4o-mini snapshot Radar used on 14 January 2027. Before switching, we measured the new model against the old one on the same questions. On our test brand, with the fixes in place, scores under the two models differed by 4 points or less. The testing also exposed four problems in our own scoring, and fixing them can raise some scores with no change on your site.
What Did OpenAI Change, and Why Did Radar Have to Move?
OpenAI's deprecations page lists five API snapshots for shutdown on 14 January 2027: gpt-4o-2024-08-06, gpt-4o-2024-11-20, gpt-4o-mini-2024-07-18, gpt-4.1-2025-04-14 and gpt-4.1-mini-2025-04-14. It recommends gpt-6-luna as the replacement. The entry is titled "2026-10-14: GPT-4 series model deprecations" (OpenAI deprecations, read 16 October 2026). Radar used the gpt-4o-mini snapshot on that list.
What did the old model do inside Radar?
Radar used gpt-4o-mini in three roles. It generated the OpenAI API answers behind the ChatGPT column in your report. It judged parts of several scores: the sentiment of an answer, which brands in an answer are competitors, and whether facts about your brand are correct. And it supplied brand and category detection, and the category shapes the questions Prompt SOV asks. As of the 16 October 2026 release, all three run on gpt-6-luna.
Why not wait until January?
We chose to switch well before the shutdown date, so we could test the replacement, fix what the testing found, and publish a dated notice with the release. Moving in October leaves three months of margin.
Radar's results describe its API measurements, not the consumer ChatGPT app. As Radar's published disclosure says, consumer apps may differ, because they can use different models, memory, personalization and live web search.
How Did We Test the New Model Before Switching?
We ran the same questions in three setups, three times each: the pre-release scoring on the old model, the fixed scoring on the old model, and the fixed scoring on gpt-6-luna. That compares the code changes while holding the model constant, then compares the models while holding the fixed code constant. Individual answers still vary between runs. The test brand was OpenAI, a public brand whose own products and website came up throughout the test answers, which is what exposed two of the counting problems.
How we compared the code changes and the model change
Same questions, three runs per setup, test brand OpenAI, 16 October 2026.
A: pre-release
Original scoring on gpt-4o-mini, the production baseline
B: fixes, old model
A to B compares original and fixed scoring on the old model
C: fixes, gpt-6-luna
B to C compares the two models on the fixed scoring
What did the matched test show?
| Check | A: pre-release (old model) | B: fixes, old model | C: fixes, gpt-6-luna |
|---|---|---|---|
| Citation Check | 99 (A) | 99 (A) | 99 (A) |
| Prompt SOV | 79 to 82 (B) | 89 to 90 (A) | 88 to 91 (A) |
| Source Influence | 76 to 77 (B) | 85 to 86 (A) | 85 (A) |
| Answer Engine | 85 (A) | 85 (A) | 85 to 89 (A) |
On this one brand, B and C differed by 4 points or less, with no grade change. From A to B, Prompt SOV went from a range of 79 to 82 to a range of 89 to 90, and Source Influence from 76 to 77 to 85 to 86, both moving from B to A. The larger differences were between the original and fixed scoring.
The judging steps were tested separately, on fixtures with a synthetic brand. Their scores were identical in all three setups: hallucination check 100, brand identity 100, a single-thread Reddit case 30 (D), a multi-thread Reddit case 55 (C), and a YouTube case 47 (D). Both models also missed the expected label on two hand-written severity cases, with the new model closer.
Prompt SOV on the test brand, by setup
Midpoint of the score range across three runs per setup. Test brand OpenAI, 16 October 2026.
What do these tests not show?
They cover one public brand and three runs per setup, so they show how large the differences were for that brand, not how every brand will move. The judging fixtures used a synthetic brand. Your own change depends on your answers, your category and your competitors.
Did the testing catch anything outside scores?
Yes. Testing our sales agent's actual requests showed that gpt-6-luna returned a server error whenever two of its settings were sent. We removed those two settings for gpt-6-luna before release, and every sales-agent test then passed.
What Did the Testing Catch in Our Own Scoring?
Testing the new model exposed four problems in how Radar scored answers, and all four are now fixed. Three were in the pre-release scoring, and comparing two models side by side made them visible. The fourth, paraphrased Reddit quotes, appeared with the new model.
Why were a brand's own products counted as competitors?
Prompt SOV measures how often a brand is named compared with the alternatives an answer recommends. The old extraction step could count the brand's own products and background technologies as competitors. For OpenAI it counted OpenAI's own products as separate rivals, and in one production run it counted WordPress. Since the 16 October 2026 release, the revised rule instructs the extraction to keep genuine alternatives in the brand's market and leave out the brand's own products, add-ons, sub-brands and background technologies.
Why did Source Influence drop the audited site?
Source Influence treats a list of large infrastructure domains, such as google.com and github.com, as background. That list included openai.com, so an OpenAI audit lost credit every time an answer cited OpenAI's own site. Since the release, the audited site is no longer filtered out, whatever list it appears on.
What went wrong with cut-off answers?
ChatGPT answers that hit the length limit were scored as if they were complete, so a score could rest on an unfinished answer. In Citation Check, Prompt SOV, Source Influence and Answer Engine, a cut-off, refused, empty or otherwise incomplete ChatGPT answer is now treated like a failed request and is never scored. Those answers also get more room, up to 1,600 output tokens.
What changed in Reddit Monitor?
When Reddit Monitor falls back to Perplexity search, the new model sometimes paraphrased the quotes it extracted. In testing, that turned a Reddit test score from D to F before the fix. The monitor now keeps a quote only when it matches the source word for word, and counts each thread once.
How Prompt SOV counts competitors now
Illustrative summary of the counting rule, not a measured result.
- The brand's own products could count as rivals
- Background technologies could count as competitors
- Cut-off answers were scored as complete
- Extra names lowered the brand share
- Extraction is instructed to leave out the brand's own products
- Extraction is instructed to keep market alternatives only
- Cut-off answers are treated as failed requests
- Share and rank use the revised competitor list
“These are results for one test brand, not an adjustment applied to every score.”
Radar scoring notice, 16 October 2026
Which Radar Scores Can This Change Affect?
As of the 16 October 2026 release, eight checks can move, because ChatGPT's answers or the model's judgment feed their scores. Six have no model input in their scores, so changing the model does not change them. Those six can still move when your site changes.
| Check | Can the model change affect it? | Why |
|---|---|---|
| Citation Check | Yes | ChatGPT answers and model-judged sentiment |
| Prompt SOV | Yes | ChatGPT answers, competitor rule, category from the model |
| Source Influence | Yes | ChatGPT answers, audited-site fix |
| Answer Engine | Yes | ChatGPT answers, incomplete-answer handling |
| Hallucination Check | Yes | The model supplies ground truth, entity checks and flag severity |
| Brand Disambiguation | Yes | The model judges whether answers describe your business |
| Reddit Monitor | Yes | Model-judged risk, verbatim quote rule |
| YouTube Monitor | Yes | Model-judged sentiment |
| AI Readiness, Crawl Check, llms.txt Validator, Robots Analyzer, AEO Auditor, Schema Audit | No | No model input in the score; only the summary text may read differently (Schema Audit makes no model call) |
The dated scoring notice lists the same eight checks.
What happened when we ran our own site after the switch?
We audited pixelmojo.io in production at 10:06 UTC on 16 October 2026, on the old model, and again at 12:21 UTC that day, after the release. Nothing on the site changed between the runs. This is a before-and-after view of the whole release, not a measure of the model's effect alone.
pixelmojo.io, same day, before and after the release
Radar production audits, 16 October 2026. Our own site, unchanged between the runs.
Overall Radar score
OpenAI probe calls using gpt-6-luna in the second audit
gpt-4o or gpt-4o-mini calls observed after the release on 16 October 2026
What Should You Do If Your Score Changed?
For a baseline under the new setup, run a fresh audit and compare it with runs made after the release at about 12:17 UTC on 16 October 2026. Saved runs keep their original scores and are not regenerated, so runs before the release and after it were measured differently.
- Read the note beside the score. Each affected score shows a dated notice titled "ChatGPT results now use gpt-6-luna". It says what changed and why.
- Keep the release out of your trends. Runs before about 12:17 UTC on 16 October 2026 used the old setup. A trend that crosses that point mixes two measurement setups.
- Check your category before you scan. The new model now supplies the category detection, and the category shapes the Prompt SOV questions. If the detected category does not describe your market, correct it in the audit setup, and compare runs that used the same category.
- Look at who you are compared against. If Prompt SOV rose, open the competitor list and check whether your own products or background technologies dropped out of it. A shorter list alone does not explain why a score changed.
- Look at the answers, not only the number. AI answers vary between runs, and these tests do not set a threshold for a meaningful change. Review the answers, the competitor list and the category before deciding whether to act.
How Does Radar Keep a Score You Can Defend When Models Change?
Radar's aim is a score you can defend: when a score moves, you should be able to find out why. A score can move because your site changed, because the AI answers changed, or because our scoring changed. A model change can affect both the answers and the model-based judgments, and providers retire models on their own schedule, so this will happen again.
“When a scoring rule changes, a score can move without anything on the site changing. Every such change is dated and explained here, and the same note appears beside the score it affected.”
Radar methodology page, scoring changelog
For this release, we measured the replacement against the old model on the same questions before switching. We compared the two models on the fixed scoring, and the original and fixed scoring on the old model. We dated the change, put the note beside all eight affected scores, and added it to the scoring changelog. For the checks with no model input, a test proves the score stays identical when the model text changes.
That is the same principle behind how Radar scores in the first place, covered in A Score You Can Defend, and behind what counts as evidence in a report, covered in the AI visibility evidence standard.
See your scores on the new model
The ChatGPT Model Change: Questions Readers Ask
Common questions about this topic, answered.
Why did Radar change the model behind its ChatGPT results?
OpenAI's deprecations page, read on 16 October 2026, schedules selected GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini API snapshots for shutdown on 14 January 2027 and recommends gpt-6-luna. The list includes the gpt-4o-mini snapshot Radar used for its ChatGPT results and for parts of its scoring, so we moved those tasks to gpt-6-luna on 16 October 2026, after testing it against the old model.
Will my Radar score change?
It can. On one public test brand, OpenAI, with the fixes in place, scores under gpt-4o-mini and gpt-6-luna differed by 4 points or less, with no grade change. With the old model held constant, the fixed scoring put Prompt SOV and Source Influence about 9 points higher on that brand. These are results for one test brand, not an adjustment applied to every score, so your own change depends on your answers.
Why did Prompt SOV go up for some brands?
The old competitor extraction could count a brand's own products, add-ons and sub-brands as competitors, and sometimes counted technologies an answer only mentioned as background, such as WordPress. Those extra names lowered the brand's share. The revised rule instructs the extraction to keep genuine alternatives in the brand's market and leave those names out, so share and rank can rise without anything changing on the site.
What changed in Source Influence?
Source Influence treats a list of large infrastructure domains, such as google.com and github.com, as background. That list also removed the audited site when it was one of those domains, so an OpenAI audit lost credit for answers that cited openai.com. The audited site is no longer filtered out.
What happens to cut-off ChatGPT answers now?
In Citation Check, Prompt SOV, Source Influence and Answer Engine, a ChatGPT answer that is cut off, refused, empty or otherwise incomplete is now treated like a failed request and is never scored. Those answers also get more room, up to 1,600 output tokens. In our matched test, 21 of 48 ChatGPT answers were cut off before the change and none after.
Are my saved Radar runs recalculated?
No. Saved runs keep their original scores and are not regenerated. The release went live at about 12:17 UTC on 16 October 2026, and runs before that used the old setup, so for a new baseline compare runs made after the release.
Which Radar checks does the model change not affect?
AI Readiness, Crawl Check, llms.txt Validator, Robots Analyzer, AEO Auditor and Schema Audit have no model input in their scores, so changing the model does not change those scores. For the first five, the model only writes their summary text, which may read differently, and a test proves their scores stay identical when that text changes. Schema Audit makes no model call at all.
Does Radar measure the ChatGPT app?
No. Radar measures the OpenAI API, now with gpt-6-luna. Radar's published disclosure notes that consumer apps may differ, because they can use different models, memory, personalization and live web search.
Where is the change explained inside Radar?
A dated notice titled "ChatGPT results now use gpt-6-luna" appears beside the eight affected scores, and the same entry is the newest item in the scoring changelog on the Radar methodology page.
Measure Again After the Release
Radar's ChatGPT results changed models because OpenAI is retiring the snapshot we used, and we measured the replacement before switching. On our test brand, the larger differences were between the original and fixed scoring, not between the two models, and a dated notice appears beside all eight affected scores. Run a fresh audit, compare runs made after the release, and read the note beside any score that moved.
