Wednesdays at 2 pm EDT we hold an AI Guild briefing. Chat with MECLABS AI if you would like to register to attend.
In the session, marketers and entrepreneurs can ask questions in the webinar chat. And we answer a few of them right here…
How do you determine when to use Claude vs MECLABS AI?
Claude is a generative AI chatbot that doesn’t have a specific purpose. Although from my personal use, it is very good with large amounts of text, more so than any other generative AI I’ve tried, lending it especially well to tasks like analyzing webinar chat logs, selecting quotes from transcripts, giving feedback on long articles, and similar tasks.
MECLABS AI (available to MECLABS AI Guild members) does have a specific purpose – headlines, value propositions, and other marketing, which makes it better for marketing specific tasks.
MECLABS is built on ChatGPT, but trained with the MECLABS methodology. Since Claude is a more general AI, it is trained on all marketing-related information, not only successful marketing practices.
I also wanted to mention, that while the questions have lately focused on AI chatbots, there are other types of generative AI, like text-to-image models. Flint McGlaughlin teaches an example in Sales Funnel Design: How to combine an “old school” journal, a web page builder, and DALL·E.
How do you overcome the 2,000-character prompt limit on Bing AI?
Here’s a fun little trick. The 2,000-character prompt limit on Bing AI only applies to the ‘more balanced’ conversation style. If you choose the ‘more creative’ or ‘more precise’ conversation style, your prompt limit doubles to 4,000 characters.
Nice, but maybe still not enough for what you want. There are two options I can see.
One, you could use Claude. Again, Claude can consume some very long prompts.
However, on the downside, Claude is a self-contained chatbot and not connected to the internet like Bing. If you need internet connection, you could always treat Bing like a customer and take a Micro-Yes Sequence approach. Map out your information (like you would a customer journey) and spoon feed it along a path for each response.
Even that is not limitless though. I’ve seen data that show Bing has a limit of anywhere from five to 20 chat turns per session. I’m not sure the exact limit, but from my experience it varies based on the topic and length. And for what it’s worth, at some point Bing just acts like it is ‘just so over’ the string of conversation and wants to start a new one. I’m not sure if it’s a feature or a bug, but it reminds me of my teenage daughter.
When you go to Claude, is that automatically Claude 2?
Yes… and no.
When you go to the Claude website, you are automatically using Claude 2, which was released in July 2023. (And as I’ve mentioned before, the most exciting thing to me about Clause is the amount of information it can process. “Claude 2 has been trained to have a further expanded context window of 200K tokens, corresponding to roughly 150,000 words” according to Model Card and Evaluations for Claude Models).
Why you even have to ask this question, and what is interesting to me as a marketer – when you visit the Claude website, it doesn’t mention a model number. This is very different from ChatGPT, which clearly displays that you are using GPT-3.5 and even shows you the ability for a paid upgrade to GPT-4.
When I ask Claude itself the reason for this, it states, “The choice aligns with Anthropic’s overall philosophy of developing AI responsibly, with a focus on user benefits rather than technology marketing. But these are just my speculations on Anthropic’s possible motivations. The company may have other reasons as well.”
If this is the case, I think it’s a bit of company logic. While ChatGPT’s use of model numbers is a marketing tactic to try to show progress and encourage upsells (just like when I bought a Samsung S23 although I would have been totally fine with the S22), it’s also a little more transparent. And with the amount of hesitancy and outright fear around AI, a little transparency is a good thing.
Can you use custom instructions on the ChatGPT version that’s in Bing?
No. As of right now, custom instruction is only for ChatGPT Plus (the paid version), and not yet available in the UK and EU. You can read more about ChatGPT customer instructions in Choosing Your AI Partner: A critical comparison of ChatGPT and Claude 2 for digital marketing.
What is Priority One if we’re going to move into these methodologies? How do we figure out where to start?
This question refers to the AI prompt methodology Flint McGlaughlin presented in this week’s briefing and LiveClass.
Like any type of conversion optimization (which this still is, even if AI is assisting you), start where you will have the most impact.
Here is an example (from Web Optimization: VacationRoost implements 2 testing methodologies to boost total conversion rates by 12%.):
Upcoming tests are elected based on a prioritization spreadsheet.
While many testing teams may allocate importance to factors such as volume and bounce rates, VacationRoost’s spreadsheet is organized by which level in the funnel the page or test would impact.
“Really what you want [to ask] is, ‘What’s going to have the most dollar impact on the organization?’” Hutchings explained.
He added that the pages leading into the purchase funnel rise to the top of the testing docket because of the likelihood of increasing conversion for the whole site is far greater on those types of pages.
I used the suggested prompts to do the competitor analysis and as my website is new, it kept referring back to the content of a website that had a similar name, even when I provided my URL. Does anyone have any other suggestions on what I can do in my prompts?
Then you may have to do this a little manually, so to speak. And here is where Claude can come in handy, since it can accept such a long prompt.
Copy and paste key information from your website (like your homepage, maybe other key pages) along with competitors’ sites into individual text files and attach them to your Claude prompt using the paper clip icon, telling Claude what you are including.
In general, answers that we get from ChatGPT, are text rich. Is that something we use as a reference or work to absorb it?
Everyone has a different working preference. But remember, the great thing about generative AI is that you don’t have to settle for the first output. If it is too long for you, tell ChatGPT what length you are looking for. Here’s an example…
I used the Prompt Kit (available to MECLABS AI Guild members) marketing prompt for audience research, and got this output from ChatGPT:
Certainly, as a marketing professional, I recognize the transformative potential of artificial intelligence (AI) in the marketing field. Yet, there are still substantial challenges in implementing AI effectively. Here’s an analysis of three major challenges, their causes, impacts, and potential solutions.
- Challenge: Data Privacy and Compliance
- Why it Exists: AI relies heavily on data for personalization and targeting. The ever-changing regulations surrounding data privacy, such as GDPR and CCPA, create complexity.
- Impact on My Life: Ensuring compliance with various regulations demands extensive legal oversight and continuous adaptation of AI algorithms, consuming both time and resources.
- Potential Solutions: Implementing strict data governance policies and utilizing AI models that prioritize data anonymization can help. Collaborating with legal experts specializing in privacy law is also essential.
- Challenge: Integrating AI with Existing Systems
- Why it Exists: Many organizations have legacy systems that aren’t built to support modern AI algorithms, leading to compatibility issues.
- Impact on My Life: Integrating AI into these systems can be labor-intensive, time-consuming, and may create bottlenecks in marketing processes.
- Potential Solutions: An incremental approach to integration, supported by a cross-functional team including IT experts, could streamline the process. Adopting flexible AI models that can be adapted to varying systems might also ease the integration.
- Challenge: Bias in AI Algorithms
- Why it Exists: AI models are trained on historical data, and if that data contains biases (gender, racial, etc.), the model may perpetuate those biases.
- Impact on My Life: Using biased AI models can lead to unethical targeting and representation in marketing, potentially damaging the brand’s reputation.
- Potential Solutions: Regular audits of AI algorithms, involving diverse teams to understand and minimize biases, and utilizing fairness-aware algorithms can help overcome this issue.
The integration of AI in marketing presents valuable opportunities but also notable challenges related to data privacy, system integration, and algorithm biases. By adopting robust legal compliance measures, embracing a cross-functional integration strategy, and rigorously auditing algorithms for bias, these challenges can be mitigated. This thoughtful approach to AI implementation aligns with the ethics and efficiency that today’s marketing professionals strive to uphold.
I then told ChatGPT, “What is the single most important thing I should know from your analysis?” To which it responded…
The single most important insight from the analysis is that while artificial intelligence presents immense opportunities for marketing, its responsible and effective implementation requires careful consideration of legal compliance, integration with existing systems, and the elimination of biases. The underlying theme is that success with AI in marketing demands a strategic approach that balances technological innovation with ethical considerations and compatibility with existing infrastructure.
Should we have already put the custom instructions into ChatGPT when we enter this prompt?
All else being equal, ChatGPT will serve you better with custom instructions because it will understand the context of your business and your customers.
But you can use any of the prompts in the Prompt Kit (available to MECLABS AI Guild members) without using custom instructions.
Again, if you want to learn more about custom instructions, you can read our last Q&A post – Choosing Your AI Partner: A critical comparison of ChatGPT and Claude 2 for digital marketing.
Would it be worth doing a few interviews with some of the people who have and haven’t purchased from our email sequence to understand the person behind the numbers?
Yes, I love that idea. That is a great example of increasing customer intimacy.
If you’re not able to go that far, you could send emails that people can reply to (in other words, not from a ‘no reply’ email address which some companies do) and make it clear in the email that replies are welcome.
Keep in mind, with this approach you will only hearing from the extremes, so take what they say with a grain of salt. But if something is truly outstanding or really horrible, you may hear about it. And you may also get a good understand of the words customer use to discuss your product or service.
For the pre-test success metric, is it important for this KPI to have large amounts of data? So for example if I wanted to use sales as my pre-test success metric but my volumes are not in thousands, is that a problem? Where is the biggest opportunity?
This is a question about A/B testing. To quickly explain a complex topic, if your results aren’t statistically significant, what you might be seeing is just random chance. An extreme example: 2 sales is a 100% increase over 1 sale, but that increase might just be random and not because of what you were testing (we explore this topic deeper in Factors Affecting Marketing Experimentation: Statistical significance in marketing, calculating sample size for marketing tests, and more).
Which begs the question – “so how many samples (sales, clicks, etc.) do I need?”
There is no one specific answer. The answer will vary based on the conversion rate difference between the control and treatment. The pre-test planning tab of the Simplified Test Protocol (available to MECLABS AI Guild members) will help you calculate that number.
So as the questioner rightly points out, it is difficult to make a number deep in your funnel, like sales, the success metric…unless you get a whole lot of sales.
As for the biggest priority, I would refer you to the above question about ‘where to start?’ However, your answer will be constrained by where you have the budget to get a statistically significant result.
We also have a free Test Planning Scenario Tool that helps you determine the optimal testing sequence.