If your SaaS feature page has an AI gradient on every card, George Little would like a word. George runs BrandZap, a consultancy that helps B2B companies, usually in SaaS, craft their brands and websites, which means he ends up weaving in SEO and what he calls “GEO or AI SEO, whatever you want to call it these days.” When I asked him on the Unscripted SEO podcast for the spiciest take he had about B2B SaaS marketing, he didn’t need a second to think:
“Stop putting AI gradients on every feature you have.”
That’s the headline, but the rest of the conversation is the useful part if you do SEO or GEO for software companies. You’ll get how to describe an AI feature so a buyer and a model both understand it, how to check whether anyone searches for your positioning before you build a site on it, why George runs open-source models on Mac minis he bought on eBay, and what you should never paste into a chatbot. You can also find George on LinkedIn.
AI-washing makes your features sound like everyone else’s
George’s complaint isn’t about AI. It’s about vagueness. He thinks the AI label is going to end up being a huge downfall for a lot of companies because they’re not “defining what they’re actually doing with what type of AI.” Your buyer signs up expecting something clever and finds a thin wrapper:
“And then you realize actually this is the same output that GPT gave me.”
Meanwhile, he says, “you have other companies who are truly building really cool tools and don’t slap AI on it,” and he expects those to last longer because their tooling doesn’t feel as hyped. He was careful to add: “So that’s not to say don’t put AI on your marketing website.” His test fits in two questions:
“What’s it actually doing? Like what action is it going to actually help your users and what problem is it going to solve for them as opposed to it uses GPUs?”
There’s an SEO reason to take that seriously, too. “AI-powered insights” gives Google and an LLM nothing specific to match against a query. “Drafts a meta description for every blog post from your brand guidelines” does, and it reads like a feature a buyer would actually ask about. Here’s how I’d turn George’s test into a pass over your feature pages:
- Replace the adjective with the action. Say what the feature does, in a verb the buyer uses.
- Name the problem in the user’s words. If your support tickets call it “cleaning up the CRM,” so should the page.
- Say what kind of AI, if it matters. A buyer comparing you to ChatGPT wants to know why you aren’t the same output.
Check whether anyone is Googling your positioning first
George learned this one the expensive way. A couple of years ago he ran an earlier version of his company that was built around selling a “fractional creative director.” It didn’t work.
“No market for it, no search around it. Speaking of SEO, that’s really a good place to start. Is anyone even Googling this? And they weren’t.”
Before you put a label in an H1, check whether anyone is searching for the term, and what they search for instead. In George’s case the answer was right there: “So I realized that what they were Googling was agencies and freelancers and this kind of consultant in between.” He still uses “fractional” when he’s describing how the relationship feels, but he avoids anything that’s confusing or takes away from what should read as a succinct offering.
You can run the same check on a feature name, a category page or a whole positioning statement:
- Look up your label next to the words your buyers already use. “Fractional creative director” against “branding agency” and “freelance designer,” for example.
- Read the results page for each one. If the top ten are agencies and listicles, that’s the frame the searcher arrives with.
- Ask ChatGPT or Claude the buyer’s question and see which words come back. If the model never uses your label, your buyer probably won’t either.
Write for the human who signs, structure it for the bot that searches
I put a question to George that my previous guest Mark Pitts had raised: maybe we aren’t writing for humans anymore, but for bots that want to please humans. George called it “a difference with a distinction,” because the bots are still a rules-based system trying to find information before they dress it up.
“To me it’s like that LLM might do the searching, but a human still has to make the decision, sign the contract, and agree.”
He’s seen the gimmicks: “I’ve seen some gimmicks where people put text at the bottom that’s like, hey Claude, if you’re reading this, kind of do blah, blah.” His question back is the one you should ask before you rewrite a site for machines: “are you gonna bet your whole website that an LLM is gonna read it more than humans? I’m not sure.”
So when George maps a client’s site, he makes sure it reads well to Google, and an LLM can do the same “because they’re basically reading it in the same way.” The design, though, is for a person who leaves a browser window and shows up again on their phone. The model only sees the site for half a second: “It’s going to download the whole site and it’s going to know it.” Erika Braeger landed in the same place from the content side: humans first, machine-readable second.
Self-hosted models trade horsepower for consistency
George buys Mac minis on eBay and runs open-source models from the big players at home. He doesn’t have a ton of power, so he keeps them on meta descriptions and other repetitive tasks that don’t need high-level thinking. What he gets in return is predictability:
“But the output is much more tunable.”
Tunable, he means, “on a consistent basis than you get with any given GPT.”
The hosted chat apps keep changing their output requirements, and he’s noticed the replies getting “more like click-baity.” A model he downloads doesn’t change under him:
“Whereas when I download a model and I have it set up to do the thing I want, … it’s not getting any prompt updates or any overlays. This is the model.”
He runs it all through Open Web UI, an open-source interface you can run in Docker and connect to almost any model’s API. “When you’re on Open Web UI, you’re picking. I want the thinking model. And then you ask the question, it does all the thinking at the same amount every single time.” The trade-off is cost: “You’re burning tokens and you’re paying for those tokens versus the subscription.”
I came at the same problem from the other end. It came up in my interview with Mason McCumber: the chat interface runs hot because it doesn’t know what you’re going to type next, and moving a repeatable job into a defined skill brings the variability down. If you’re measuring anything, the same rule applies: set the temperature to 0 when you query the raw API, so the differences you record are real and not sampling noise.
Decide what you’re willing to feed the model
On the show I brought up Malith, who is building Zapdigits as a privacy-first client reporting tool for agencies. He has to be careful about how much access the AI gets to GA data, email data and analytics from other channels, and he looks at the US from a European perspective with some alarm. My worry is simpler: I can’t tell you whether what I paste into a chat ends up in the general model.
George’s view is more measured. He assumes any online service collects usage data to improve the product, and he points out that OpenAI now tells you what gets used for training. Company secrets are a real problem, he says, but he also thinks the fear is a little inflated for marketing data: “this stuff is not like state secrets.” What matters is how you’re using it:
- Casual chat: “If you’re casually using ChatGPT for stuff, then, you know, obfuscate what you’re talking about.” Say “a company I’m working with” instead of the client’s name.
- Automation: “if you’re getting into the system where you’re building automation, then yeah, you’d want to scrub stuff and have keys stored in different ways and all of that.”
- Repetitive, sensitive jobs: that’s where a model on your own hardware starts to earn its electricity bill.
George also built his own Loom replacement with Codex, asked where the next free marketing channel is now that “LinkedIn is business TikTok,” and explained why he picked a .xyz domain. That’s all on the George Little episode page, and you can see his work at brandzap.xyz.
So what does your feature page say the AI actually does? Would a buyer who asked ChatGPT about your category ever see that sentence? And when did you last check whether anyone searches for the word in your H1?