Evidence-Based AI: First Search, Then Research

Published on November 4, 2025

AI Search Gets Closer to Getting It Right.

The basic uncertainty and fear surrounding artificial intelligence involves the jobs that generative AI might replace. However, PR practitioners find that people must be involved in GenAI research. Search engine/AI hybrids can provide more trustworthy results but cannot replace the human element. 

Search-augmented AI tools such as Perplexity and Google AI Overviews crawl the internet frequently, which makes their results more current and their PR ideas more pitch-worthy. But the real strength of search-supplemented AI is the trail it leaves for researchers to follow.

Search-oriented AI labels its sources with text, not an inconspicuous chain icon. These descriptions give market researchers a better idea of how fresh and thorough the underlying content is and where errors in the summary might lie. The links are only a start, though. Users should read the descriptions and click through to dive into the source material. 

Computer brains are by no means know-it-alls. AI has a limited knowledge base. Its language processors mostly interpret user questions and summarize search results, without much analysis. Still, AI is an incredible resource for finding things, even when incredible turns out to mean “not at all credible.” 

What Mama Said: Fact-Checking GenAI

Ex-journalists on the Purpose Brand team were schooled in the Chicago newsroom adage, “If your mother says she loves you, check it out.” Listening to eyewitness stories and publicity pitches has fine-tuned their BS detectors, which have proved equally useful for AI fabrications. 

AI can turn a firehose of information into a powerful stream, but the user still has to grab hold and direct the flow. With anything short of common knowledge, marketers should prompt a chatbot to give citations and check them thoroughly. 

GenAI is not a final product, but a spray of pure fact, murky assertion and a dense bit of creative writing. The PR fact-checker can filter out the toxic bits, and the process involves the kinds of things researchers normally do:

• Go to the Go-To Source. When you need specifics, GenAI might be generalizing. In one close-to-home case, an AI-generated 2025 summer reading list described fake books convincingly enough to fool the writer, editor and Chicago client of King Features Syndicate. An interested bibliophile likely would have checked the title on Amazon, and an editor doing the same would have revealed the books were fabrications similar to previously published titles. 

• Consider the Source. AI can pick up cues from context but may not be able to weigh the author or publication’s reputation. Is the writer a journalist, an advocate, an influencer or a pseudonym? Are their platforms well established or freshly minted? How are they funded? 
• Let’s Do the Time Warp Again. GenAI often does not have access to the most current information and may be captive to its training data, citing stale reporting or old calendar dates.
• Bateman vs. Batman. Lacking firsthand knowledge, AI may draw inferences from related topics. Reviewing the citation will indicate how relevant the source material is to the subject at hand.

• What’s the Motivation? There’s a difference between fact and “based on a true story.” Fake news has been with us for centuries, whether for dramatic or deceptive effect. Do the writers seem to inform, persuade, entertain, sell a product, or spread propaganda? Do they explain the topic or just make assertions?

It won’t take long to get a sense of how far to trust either the AI or the source material, but frankly, either can trip up researchers. Internet sources could be firsthand reports or several steps removed from boots-on-the-ground witnesses. AI processing is more likely than humans to misread context, but either can misinterpret the facts or take them at face value. 

Statistics can mislead. The tragic example is the claim that opioid patients do not develop drug dependency. Oxycodone advertising cited a general comment on addiction treatment as if it were a peer reviewed, published finding about the drug. But even clinical trials can be overgeneralized. Early studies don’t always pan out, as when chloroquine failed to meet expectations as a COVID-19 treatment. It’s a recurring issue when clinicians try to apply the results from a small study to larger groups or from a general to a minority population. 

In disciplines from self-help to high-tech, fact-checkers see a range of ballpark guesses, projections and test runs repeated over time as proven facts. Search engines often give these statements great weight, and chatbots accept them as conventional wisdom. PR practitioners and their clients should not amplify these errors. Nothing is, as newsroom sarcasm puts it, “too good to check.”

It may not seem reassuring that GenAI will require researchers to dive down rabbit holes. Media relations professionals might wonder if the process saves them any time. If it’s any consolation, they’ll emerge from their deep dive better informed about the topic and more persuasive about its nuances.

How Do You Feel About Sentiment Analysis?

One attraction to generative AI is that it can sift through and generalize high-volume information sources such as media mentions or survey responses. Even here, though, GenAI has limited capacity for data analysis. Chatbots can misread sentiment, especially when they use a model trained on formal writing instead of slangy or sarcastic customer feedback. When numbers are involved, an AI process is not designed to “do the math” or even compare proportions. 

Still, AI summaries can point in the right direction to monitor brand reputation, improve call-center response or identify a consumer need. Marketers can gear their data governance practices to refine brand positioning or monitor media mentions, social media posts and product reviews:

• Make Comments Safe for AI. Use a data cleaning tool to filter out personal information, spam and irrelevant code, or to translate acronyms or emojis. Control who can see the raw data as well as the analysis.
• Compare to Benchmarks. If possible, bring in accurate data from other sources to give GenAI a firmer foundation for behavior patterns. Compare social media mentions with sales data, website conversions or call center logs. Know whether the data sources adequately represent minority or niche target audiences.
• Watch for Blind Spots. Try to distinguish aspects that drive positive or negative comments. If the goal is to raise the brand’s profile, take a “neutral” media mention as a win. Social media complaints may have only a limited range or short-term effect. 

Automating data-intensive or rote tasks can free up time for strategic insight and creative practice. Chatbots help marketers pursue new audiences and be more effective in engaging them and improving lives. By casting a wider information net, informed communicators catch the details that matter to their audiences.