SlideShare a Scribd company logo
Copywriting with AI
How creative can machines
really be?
Wijnand Meijer – Co-founder & CEO
@wijnandmeijer & linkedin.com/in/wijnandmeijer
Weren’t humans supposed to write?
Introduction of GPT-3
Overview of writing assistants using GPT-3
How I used Jarvis.ai
Performance of Jarvis headlines so far
Closing thoughts
Agenda
1
2
3
4
5
6
Weren’t humans
supposed to write?
The holy grail of marketing
It’s all about reaching….
The right person with the right message at the right time (and place)
• Keywords  Feeds, Shopping,
DSA
• Demographics
• 1st party audiences + look-a-likes
• Other audiences (Affinity, In-
market, Employment, Education)
 Smart campaigns / Smart bidding
• STA  ETA  RSA
• Manual A/B testing  “Optimize”
 RSA asset impressions / Ad
Variations
• Text ads, display ads, videos,
landing pages, any content.
• Human-generated 
AI-assisted / AI – generated?
Manual CPC  Bid adjustments 
Smart bidding
(targeting) (creative) (bidding)
The message reigns again (back to Mad Men)
• Right person and right time is slowly getting taken away from us through competition, algorithms or simply by being
priced out of some of these areas. When all media becomes mass media, it’s the message that matters. We have to
become copywriters… again.
• If you’re only good at targeting, you’re a mechanic, not a marketer. Marketing is the crafting and the amplification of
a message. The last 15 years, it’s all been about amplification, and it could work with a ‘meh’ message.
• The coming years, the edge is going to be on the messaging side. You’ll have to be a communicator first.
The amplification of that message is table stakes, everyone (or every algorithm) will know how to do that.
Ryan Deiss – Perpetual Traffic episode 334 (must-listen!)
So what should we do?
A: Become the best human copywriter we can possibly be? B: Let the AI take over and find a new job?
Or C: something else?
Introduction of GPT-3
Kevin Lacker - Giving GPT-3 a Turing Test
Introduction of GPT-3
• GPT-3 = Generative Pre-trained Transformer
3
• Generative means that the model’s goal is to
generate text. It wants to predict what word
comes next.
• Pre-trained means that the algorithm doesn’t
need any more training. It has “read”
3,375,000,000 pages of text, including all of
Wikipedia.
• Transformer is the algorithm the model uses.
It specializes in how words are used in natural
language.
• In short: it’s autocomplete on steroids
Introduction of GPT-3
Introduction of GPT-3
https://lifearchitect.ai/models/
Overview of writing
assistants using GPT-3
Overview of writing assistants using GPT-3
Anyword Copy.ai Copysmith Jarvis.ai
Unbounce
Smart Copy
Writesonic
Rytr Scalenut
You’ll find even more on gpt3demo.com and the AI Writing Assistants category on G2
Have fun with some less serious applications
Categorized conversation-starters
How I used Jarvis.ai
• So far, it seems to be the best-
reviewed and best documented AI
writing assistant.
• It also already acquired two other
strong players in this category:
Headlime and Shortly.
• Their resources include:
– Bootcamp (start here)
– Help center
– Academy (webinars)
– YouTube channel
– Facebook group with 42K members
– Integrates with Surfer SEO
– Just released an API (beta waitlist)
Why I picked Jarvis.ai
Jarvis templates: 47 and counting
Tone of voice: endless possibilities
Absurd
Accusatory
Acerbic
Admiring
Aggressive
Aggrieved
Ambivalent
Amused
Angry
Animated
Apathetic
Apologetic
Appreciative
Ardent
Arrogant
Assertive
Awestruck
Belligerent
Benevolent
Bitter
Callous
Candid
Caustic
Cautionary
Celebratory
Chatty
Colloquial
Comic
Compassionate
Complex
Compliant
Concerned
Conciliatory
Condescending
Confused
Contemptuous
Critical
Cruel
Curious
Cynical
Defensive
Defiant
Demeaning
Depressing
Derisive
Detached
Dignified
Diplomatic
Disapproving
Disheartening
Disparaging
Direct
Disappointed
Dispassionate
Distressing
Docile
Earnest
Egotistical
Empathetic
Encouraging
Enthusiastic
Evasive
Excited
Facetious
Farcical
Flippant
Forceful
Formal
Frank
Frustrated
Gentle
Ghoulish
Grim
Gullible
Humble
Humorous
Hypercritical
Impartial
Impassioned
Imploring
Impressionable
Inane
Incensed
Incredulous
Indignant
Informative
Inspirational
Intense
Intimate
Ironic
Irreverent
Jaded
Joyful
Judgmental
Laudatory
Light-Hearted
Loving
Macabre
Malicious
Mean-Spirited
Mocking
Mourning
Naïve
Narcissistic
Nasty
Negative
Nostalgic
Objective
Obsequious
Optimistic
Outraged
Outspoken
Pathetic
Patronising
Pensive
Persuasive
Pessimistic
Philosophical
Playful
Pragmatic
Pretentious
Regretful
Resentful
Resigned
Restrained
Reverent
Righteous
Satirical
Sarcastic
Scathing
Scornful
Sensationalistic
Sentimental
Sincere
Sceptical
Solemn
Subjective
Submissive
Sulking
Sympathetic
Thoughtful
Tolerant
Tragic
Unassuming
Uneasy
Urgent
Vindictive
Virtuous
Whimsical
Witty
Wonder
World-Weary
Worried
Wretched
Or any celebrity
with a distinctive
tone of voice, like:
Donald Trump
Elon Musk
Kanye West
Oprah
Tony Robbins
Jarvis warming-up: Content Improver template
Jarvis Perfect Headline template (TrueClicks input)
Jarvis Google Ads Headline template (TrueClicks input)
42 characters
20 characters
67 characters
43 characters
38 characters
Jarvis Perfect Headline template (external input)
Performance of Jarvis
headlines so far
• Original plan: rotate two ETAs in each ad group “Human vs AI” fight till the…
CTR significance
• But then I remembered Martin Röttgerding’s “Debunking Ad Testing”
(slides + part 1, part 2 part 3) talk from 2017 and quickly dropped that plan.
• Next plan was to use Ad Variations and A/B test Human vs AI headlines
across many RSAs….
• But then I realized almost none of the headlines I eventually used were
100% Jarvis-generated, I almost always edit them a bit.
• Eventually, I simply added lots of new RSAs containing a mix of fully
human-written headlines and new “AI-assisted” headlines and let them run
for a while, to see which ones will get the most impressions.
• So this test has about 0% scientific value, but my goal for now is just to find
out if it’s promising enough to keep using it.
My “testing methodology”
AI-assisted vs 100% human impressions so far
Closing thoughts
GPT-3’s strengths & weaknesses
Headlines
Strengths
• GPT-3 is great for short and creative content like:
• Headlines, titles, tweets
• Product descriptions
• Paragraphs
• It can generate dozens of these in seconds, which is
great for brainstorming and inspiration. No more
staring at the blank page.
• It’s not plagiarism. It’s “inspired by” human text, but not
copying it.
Weaknesses
• Can’t write a narrative thread, with coherence and logic.
• Makes up stuff (“fake news”). Makes claims without
evidence.
• Can’t generate real new insights or new perspectives.
• Contains (human-based) biases around religion, gender,
etc.
Be sure to read GPT-3 is the Sparring Partner You Didn’t Know You Needed and
The Search Singularity: How to Win in the Era of Infinite Content
So what should you do?
1. Know how to use AI for the stuff it’s good at
• Getting good output is more art than science, keep tweaking the inputs, templates, tone of voice, until
you’re getting good outputs.
• Don’t copy-paste any result without (fact-)checking and making sure it matches your tone.
• It’s definitely worth the $20-$100/mo for anyone in marketing. Whether you’re doing PPC, SEO, social,
content… we all need to write (even if it’s just emails) and this will speed it up.
• In short: use AI-assisted text when it makes sense, but don’t publish 100% AI-generated text.
2. Know how to create content only humans can
• Still read those 6 copywriting/psychology books from the
start, applying what’s in there will put you in the top 1%.
• Create content around the five sources of thought leadership,
this ensures it will be unique, human, and impossible to
copy.
Creating thought leadership content goes beyond PPC
or search, but is crucial for any brand or agency to stay
relevant and noticeable.

More Related Content

Copywriting with AI - How creatives can machines really be?

  • 1. Copywriting with AI How creative can machines really be? Wijnand Meijer – Co-founder & CEO @wijnandmeijer & linkedin.com/in/wijnandmeijer
  • 2. Weren’t humans supposed to write? Introduction of GPT-3 Overview of writing assistants using GPT-3 How I used Jarvis.ai Performance of Jarvis headlines so far Closing thoughts Agenda 1 2 3 4 5 6
  • 4. The holy grail of marketing It’s all about reaching…. The right person with the right message at the right time (and place) • Keywords  Feeds, Shopping, DSA • Demographics • 1st party audiences + look-a-likes • Other audiences (Affinity, In- market, Employment, Education)  Smart campaigns / Smart bidding • STA  ETA  RSA • Manual A/B testing  “Optimize”  RSA asset impressions / Ad Variations • Text ads, display ads, videos, landing pages, any content. • Human-generated  AI-assisted / AI – generated? Manual CPC  Bid adjustments  Smart bidding (targeting) (creative) (bidding)
  • 5. The message reigns again (back to Mad Men) • Right person and right time is slowly getting taken away from us through competition, algorithms or simply by being priced out of some of these areas. When all media becomes mass media, it’s the message that matters. We have to become copywriters… again. • If you’re only good at targeting, you’re a mechanic, not a marketer. Marketing is the crafting and the amplification of a message. The last 15 years, it’s all been about amplification, and it could work with a ‘meh’ message. • The coming years, the edge is going to be on the messaging side. You’ll have to be a communicator first. The amplification of that message is table stakes, everyone (or every algorithm) will know how to do that. Ryan Deiss – Perpetual Traffic episode 334 (must-listen!)
  • 6. So what should we do? A: Become the best human copywriter we can possibly be? B: Let the AI take over and find a new job? Or C: something else?
  • 7. Introduction of GPT-3 Kevin Lacker - Giving GPT-3 a Turing Test
  • 9. • GPT-3 = Generative Pre-trained Transformer 3 • Generative means that the model’s goal is to generate text. It wants to predict what word comes next. • Pre-trained means that the algorithm doesn’t need any more training. It has “read” 3,375,000,000 pages of text, including all of Wikipedia. • Transformer is the algorithm the model uses. It specializes in how words are used in natural language. • In short: it’s autocomplete on steroids Introduction of GPT-3
  • 12. Overview of writing assistants using GPT-3 Anyword Copy.ai Copysmith Jarvis.ai Unbounce Smart Copy Writesonic Rytr Scalenut You’ll find even more on gpt3demo.com and the AI Writing Assistants category on G2
  • 13. Have fun with some less serious applications
  • 15. How I used Jarvis.ai
  • 16. • So far, it seems to be the best- reviewed and best documented AI writing assistant. • It also already acquired two other strong players in this category: Headlime and Shortly. • Their resources include: – Bootcamp (start here) – Help center – Academy (webinars) – YouTube channel – Facebook group with 42K members – Integrates with Surfer SEO – Just released an API (beta waitlist) Why I picked Jarvis.ai
  • 17. Jarvis templates: 47 and counting
  • 18. Tone of voice: endless possibilities Absurd Accusatory Acerbic Admiring Aggressive Aggrieved Ambivalent Amused Angry Animated Apathetic Apologetic Appreciative Ardent Arrogant Assertive Awestruck Belligerent Benevolent Bitter Callous Candid Caustic Cautionary Celebratory Chatty Colloquial Comic Compassionate Complex Compliant Concerned Conciliatory Condescending Confused Contemptuous Critical Cruel Curious Cynical Defensive Defiant Demeaning Depressing Derisive Detached Dignified Diplomatic Disapproving Disheartening Disparaging Direct Disappointed Dispassionate Distressing Docile Earnest Egotistical Empathetic Encouraging Enthusiastic Evasive Excited Facetious Farcical Flippant Forceful Formal Frank Frustrated Gentle Ghoulish Grim Gullible Humble Humorous Hypercritical Impartial Impassioned Imploring Impressionable Inane Incensed Incredulous Indignant Informative Inspirational Intense Intimate Ironic Irreverent Jaded Joyful Judgmental Laudatory Light-Hearted Loving Macabre Malicious Mean-Spirited Mocking Mourning Naïve Narcissistic Nasty Negative Nostalgic Objective Obsequious Optimistic Outraged Outspoken Pathetic Patronising Pensive Persuasive Pessimistic Philosophical Playful Pragmatic Pretentious Regretful Resentful Resigned Restrained Reverent Righteous Satirical Sarcastic Scathing Scornful Sensationalistic Sentimental Sincere Sceptical Solemn Subjective Submissive Sulking Sympathetic Thoughtful Tolerant Tragic Unassuming Uneasy Urgent Vindictive Virtuous Whimsical Witty Wonder World-Weary Worried Wretched Or any celebrity with a distinctive tone of voice, like: Donald Trump Elon Musk Kanye West Oprah Tony Robbins
  • 19. Jarvis warming-up: Content Improver template
  • 20. Jarvis Perfect Headline template (TrueClicks input)
  • 21. Jarvis Google Ads Headline template (TrueClicks input) 42 characters 20 characters 67 characters 43 characters 38 characters
  • 22. Jarvis Perfect Headline template (external input)
  • 24. • Original plan: rotate two ETAs in each ad group “Human vs AI” fight till the… CTR significance • But then I remembered Martin Röttgerding’s “Debunking Ad Testing” (slides + part 1, part 2 part 3) talk from 2017 and quickly dropped that plan. • Next plan was to use Ad Variations and A/B test Human vs AI headlines across many RSAs…. • But then I realized almost none of the headlines I eventually used were 100% Jarvis-generated, I almost always edit them a bit. • Eventually, I simply added lots of new RSAs containing a mix of fully human-written headlines and new “AI-assisted” headlines and let them run for a while, to see which ones will get the most impressions. • So this test has about 0% scientific value, but my goal for now is just to find out if it’s promising enough to keep using it. My “testing methodology”
  • 25. AI-assisted vs 100% human impressions so far
  • 27. GPT-3’s strengths & weaknesses Headlines Strengths • GPT-3 is great for short and creative content like: • Headlines, titles, tweets • Product descriptions • Paragraphs • It can generate dozens of these in seconds, which is great for brainstorming and inspiration. No more staring at the blank page. • It’s not plagiarism. It’s “inspired by” human text, but not copying it. Weaknesses • Can’t write a narrative thread, with coherence and logic. • Makes up stuff (“fake news”). Makes claims without evidence. • Can’t generate real new insights or new perspectives. • Contains (human-based) biases around religion, gender, etc. Be sure to read GPT-3 is the Sparring Partner You Didn’t Know You Needed and The Search Singularity: How to Win in the Era of Infinite Content
  • 28. So what should you do? 1. Know how to use AI for the stuff it’s good at • Getting good output is more art than science, keep tweaking the inputs, templates, tone of voice, until you’re getting good outputs. • Don’t copy-paste any result without (fact-)checking and making sure it matches your tone. • It’s definitely worth the $20-$100/mo for anyone in marketing. Whether you’re doing PPC, SEO, social, content… we all need to write (even if it’s just emails) and this will speed it up. • In short: use AI-assisted text when it makes sense, but don’t publish 100% AI-generated text. 2. Know how to create content only humans can • Still read those 6 copywriting/psychology books from the start, applying what’s in there will put you in the top 1%. • Create content around the five sources of thought leadership, this ensures it will be unique, human, and impossible to copy. Creating thought leadership content goes beyond PPC or search, but is crucial for any brand or agency to stay relevant and noticeable.