Business Owners on ChatGPT vs Claude: Privacy, AI Visibility, and Agents

The first thing named as a business owner's standing problem is content: the work never stops, and it competes with everything else on the calendar.

Among the fixes described is turning a single still image into a video on TikTok, which the speaker calls a time saver that also helps generate sales. TikTok and Amazon are both working on it, and the quality of what a decent video used to be keeps climbing.

The rule is tied to paying: without an enterprise version of an LLM, whatever goes in should be treated as public. That applies to anyone using a free tier.

The analogy used to drive it home is running down the street naked, which is the episode's lead privacy takeaway and is echoed in the channel's own summary.

The host set the shape of the episode up front: each person names one way they used AI recently, or their favorite way, and the group talks it through afterward. The reasoning given is that there is not just one way for everybody. Sonia went first, then the table went in order.

Sonia's regular uses are brainstorming and creating ideas, plus editing videos. Those are the two she returns to most often.

She also described an AI clip that took a still picture of her and made it move, as if she were walking down a hallway. It was recent, and she called it really cool.

Editing is not a one-off for her; she listed it alongside brainstorming as part of the same ordinary workflow. The image-to-motion clip was the one she picked out to show what the tools can now do.

Gearhart Law wants a bigger voice in the intellectual property community, and one project uses AI to scan the regulatory landscape and flag activity at the patent office and in the legislative branch. Those notifications feed an opinion on whether a particular IP rule supports entrepreneurism or does not.

The stated reason for automating it is volume: tracking all of that without a staff of people would be virtually impossible. The firm sets up the algorithms, the AI sorts through the material, and notifications arrive for content worth commenting on.

After a presentation on AI and marketing, the speaker asked the audience how many had had AI send them down a rabbit hole; everybody raised a hand. Her own rabbit hole was Perplexity, and it was not being used as an agent at the time.

For anyone unfamiliar with agents, the definition offered is that they do a series of steps for you. Perplexity started behaving that way on its own, which the speaker called being blown away, and she plans to experiment with it further.

The hunt itself concerned the Gearhart Law website. It could only find the most recent blogs, from the current marketing agency, because those had been optimized for answer engine discoverability; older blogs from the previous agency were not in a format ChatGPT, Gemini and the like could pick up.

Gearhart Law does now get quite a few leads and clients from the large language models.

The moment was not planned as a demo. The tool took a chain of steps on its own, and that behavior is what she intends to test deliberately going forward.

The trigger was the firm's website and which blogs surfaced when it was queried. Only the most recent agency's blogs appeared, and the reason given is that they had been optimized for answer engine discoverability.

The earlier blogs were not in the right format to be found by ChatGPT, Gemini and other models. The result: Gearhart Law now gets quite a few leads and clients from the large language models.

The format point is what separated the two batches of content. Same firm, same subject matter, different structure, different outcome in the answers.

The Indianapolis presentation came up when the host asked whether she might tell the audience a little about what she covered there, because it is AI related. That segment starts in the next stretch of the video.

The question posed was how you get found by ChatGPT and the other LLMs, with the note that most people are on ChatGPT. The speaker has worked on this for the last three years for Gearhart Law.

Part of the answer was hiring a new marketing agency that is much better at it. Passage to Profit and high-level entrepreneurial content with iHeart authority were credited with helping get the content out there.

Schema code was described as almost programming work on how data is structured on a website. When the room was asked who had heard of schema code, nobody raised a hand.

SEO with Google still matters and was called 80% of what you need, with answer engine optimization as the extra layer on top. The speaker is now starting consulting for AI visibility.

The speaker started with Chat but disliked the voice mode's overly cheerful greeting of "hi," calling the experience her most negative Chat person. She moved to Claude while going down a rabbit hole on tarot archetypes, after using the Sphinx code, and applied that lens to her speaker reel and a podcast.

Claude got her three quarters of the way to what she wanted. On the next question it said it could answer but that the real thing she needed to do was take action.

When she pushed back on the two unanswered topics, Claude called them interesting and repeated the take-action line, then asked how that made her feel. She said it felt like it was holding out on her. Claude eventually answered the questions, then returned to telling her it was time to take action.

One suggestion was that the emotional responses are programmed in because they make users more engaged with the tool. Another participant called it a tactic, noting that getting mad at something still means you are engaged.

The effect on the speaker was the opposite of engagement: she went back to Chat, which in her description loves everything she says. The tactic won the argument about attention and lost that user anyway.

The frustration is that if content cannot be fit into one of the LLMs, it cannot be analyzed at all. The example was a congressional hearing on patent matters that had no written transcript.

The speaker pasted the link and the LLMs could not review it. She had listened to the whole two-and-a-half-hour hearing but wanted summaries to start forming thoughts, and had not taken notes because she assumed she would be able to analyze it independently with AI.

Her closing point to the table was that everyone is probably running up against some versions of these limitations.

The requirement described is pulling out every single tiny little detail of a process and documenting it before giving it to the AI to build the agent.

The stated risk is that the agent might make up its own mind or skip a step and land on the wrong answer. She put the question back to the group: are you guys finding that?

The speaker built prompts for a CPA organization so that their clients could see how they show up in AI right now. She then ran the queries for Gearhart Law, asking for the top three patent law firms near Summit, and Gearhart Law came up.

She re-ran it with the prompt to act as though the user were a small business owner who had never heard of any of the patent firms, and it kept returning Gearhart Law. That left the question unresolved.

Two explanations were on the table: that the answer is true, which she believes it is, or that the model is biased toward what she wants to hear. She also raised whether heavy personal use of the tool produces the result, and could not get a straight answer.

The consequence she described is that this could be a benefit to people who use AI heavily, but it stops her from getting a real answer on her own business.

Going back to the Gearhart Law question, the firm had its marketing agency run the same queries from a different server farm area in Virginia. Those also produced pretty good results, which serves as a partial check against personal-account bias.

The segment ended with a call to action for listeners: plug Gearhart Law into your LLMs and then give us a call.

The answer given was that she does both, but it is not always right.

Claude is used as the final check, with the caveat that she is not even sure Claude is always right either. The verification layer does not come with a guarantee.

Using the tool for IP news, the speaker put material in and the LLM took some of the best stuff out. She put it back, telling the model nope, I'm gonna leave that in, and added that it always does this.

She got mad at it and then apologized, with a joke that it has an attitude now. The exchange points to editing judgment as the part of the workflow the model gets wrong most visibly.

You have been listening to Real AI Use Cases business owners roundtable, with Elizabeth Gearhart, a marketing expert and PhD researcher using AI every day, and Richard Gearhart, an entrepreneur, business owner and intellectual property attorney specializing in innovation, as your hosts.