Cookie Consent by Free Privacy Policy Generator

Busting the Myth: How Persona Affects AI Accuracy

Go to the Best Of the SEO Community.

Mike Sonders
Mike Sonders
Dec 9, 2025, 4:32 PM
fun new research from wharton: https://papers.ssrn.com/sol3/papers.cfm?abstract<emid=5879722>

it turns out adding an expert persona to your prompt (“you are a senior software engineer”) generally has no significant impact on accuracy.

in fact, they even found that having a “physics expert” persona answer law questions vs. having a “law expert” do it made little difference. :upsidedown_face:

the more you know!
???? 6 ???? 1
Katie Stone
Katie Stone
Dec 9, 2025, 4:35 PM
That’s so interesting! One time I prompted Gemini with, “You are a Content Marketer” and it responded, “I am not a content marketer, I am a large language model,” and ever since then, I’ve stopped using the expert personas in prompts haha
:joy: 7
Kane Jamison
Kane Jamison
Dec 9, 2025, 4:36 PM
ah thank gosh. i've been waiting for this practice to go away.
???? 1
Mike Sonders
Mike Sonders
Dec 9, 2025, 4:45 PM
@user that’s so funny — i’ve never had the ai, be like, “uh, no i’m not.”
Katie Stone
Katie Stone
Dec 9, 2025, 4:46 PM
It seriously took me aback, I wish I had taken a screenshot! I haven’t been able to replicate it since :woman-shrugging:
Dorron
Dorron
Dec 9, 2025, 7:21 PM
@user Good to see actual documentation of this. I have always said the "Role" is personification for style not accuracy. It defines the framing not the outcome.
???? 2
Mike Sonders
Mike Sonders
Dec 9, 2025, 7:44 PM
@user and it could be useful for framing! (as they discuss in the paper.) just not for “accuracy”
:100: 1
Roger Marquez
Roger Marquez
Dec 9, 2025, 7:48 PM
@user Does the research says anything about ending the prompt with a "Thanks"? Does it influence result/output quality? :stuckouttongue: ????
Mike Sonders
Mike Sonders
Dec 9, 2025, 7:54 PM
@user actually there’s separate research (i’d have to find it) that indicated that results from AI are better when you’re rude vs. polite :upsidedownface:
Mike Sonders
Mike Sonders
Dec 9, 2025, 7:56 PM
https://virtualizationreview.com/articles/2025/10/28/rude-prompts-yield-smarter-chatgpt-answers-study-finds.aspx
Roger Marquez
Roger Marquez
Dec 9, 2025, 7:56 PM
```...that results from AI are better when you're rude vs. polite...```
@user Oh my God! and I thought to joke about it and that´s an actual thing!

You would blast off laughing if you saw my Claude chat when I don´t get Sonnet to follow instructions :stuckouttongue: ????

Now it makes sense, these little AI rebels.
Mike Sonders
Mike Sonders
Dec 9, 2025, 7:57 PM
there’s a bunch of caveats around this (only tested on one model, only tested one type of scenario (multiple choice questions), …)
Roger Marquez
Roger Marquez
Dec 9, 2025, 7:57 PM
```...Rude Prompts Yield Smarter ChatGPT Answers...```
Oh wait, this can actually work in my favor ????

I once even told Sonnet something like "if you don´t follow these instructions I will prompt this in Gemini" ????
Mike Sonders
Mike Sonders
Dec 9, 2025, 7:57 PM
i’ve definitely seen threatening work! (e.g., “if you don’t get this right i’ll be fired”) but that was anecdotal experience
Victor M Pan
Victor M Pan
Dec 9, 2025, 8:24 PM
????
:heavy_check_mark: 1
Victor M Pan
Victor M Pan
Dec 9, 2025, 8:27 PM
Implications on prompt tracking by substituting the role as a persona is significant
Victor M Pan
Victor M Pan
Dec 9, 2025, 8:29 PM
I do think there's directional value in subjective evaluations (best for... what?) but at the same time feelings aren't valid data points :joy: Look forward to more stuff like this
:handshake: 2
Mike Sonders
Mike Sonders
Dec 9, 2025, 8:35 PM
yeah 2.0-flash was the outlier that made me say “generally has no significant impact on accuracy”. goes to show that things are always more nuanced than the headline would lead you to believe. :slightlysmilingface:
Mike Sonders
Mike Sonders
Dec 9, 2025, 8:37 PM
> Implications on prompt tracking by substituting the role as a persona is significant
maybe not? it’s probably fine if the intention is just to see if/how the extra context (i.e., persona) changes the general content of the ai’s response, rather than testing the accuracy of the response.
Mike Sonders
Mike Sonders
Dec 9, 2025, 8:38 PM
i.e., the answer is going to (expectedly) change if i say “i’m a small biz owner looking for a crm” vs. “i’m an enterprise sales leader looking for a crm.” i think the use case being addressed by the paper is when people do the “you’re a brilliant astro-physicist” thing when they’re trying to get a “smarter” answer about physics, for example.
Victor M Pan
Victor M Pan
Dec 9, 2025, 8:50 PM
Yeah the use case is different (accuracy on test questions) but I do wonder about the issue of anthropomorphization and how eliminating role might help with that.

Channels

  • agency (3)
  • ai-for-seo (11)
  • ai-llms (24)
  • analytics (13)
  • biz-of-seo (4)
  • career (4)
  • content (15)
  • cro (4)
  • data-viz-reporting (4)
  • datascience (2)
  • ecommerce (3)
  • events (1)
  • freelance (2)
  • general-seo (21)
  • in-house (2)
  • local-seo (12)
  • local-seo,biz-of-seo (1)
  • lounge (1)
  • reading-list-resources (3)
  • search-console (11)
  • shameless-plug (1)
  • technical-seo (37)
  • tools (1)


    • Join The Community

    The SEO Community is a free organization, with private Slack group and events, for those interested in search and AI optimization to learn, discuss, and meet others in the industry.

    Powered by MODX

    Resources

    The Community Story SEO Articles SEO Jobs SEO Conferences Top SEO Discussions Newsletter Free Tools Learn Claude Code

    The SEO Community © 2026

    Shop Merch Code of Conduct Partnerships Sitemap Update cookies preferences Terms of Service Privacy Policy