AI for Research and Ideation: Audience Mining, Competitor Teardowns, Positioning Angles, and the Content Brief That Saves the Draft

Today we're turning AI into a research and ideation engine. By the end of this episode you'll be able to point it at your audience, your competitors, and your positioning, and then land all of that work in a single document that makes your next draft dramatically better. That document is the content brief, and it's the hero of this whole episode.

But before any of that, I want to give you the one idea that survives a tool swap. If you forget every prompt I read you today, hold onto this. AI is genuinely fast at two things: synthesizing a pile of information, and structuring it into something usable. What it is not, is informed. It has no idea what your real customers actually said. None. So the quality of your research output is capped, hard, by the quality of the real artifacts you feed it. Garbage in, or worse, nothing in, gives you confident fiction out.

Say that back to yourself. Confident fiction. That's the failure mode for this entire category of work, and it's seductive because the fiction is well written. It reads like research. It just isn't.

So here's the move, and it's the same move for all four jobs we're going to cover. You gather real source material. You have AI structure and synthesize it. You verify what it gave you. And you turn it into a brief. Gather, structure, verify, brief. That rhythm repeats for audience research, for competitor teardowns, for positioning, and for the brief itself. Four jobs, one move.

Let's start with job one.

Audience and customer research

The first job is understanding who you're actually talking to, and doing it without making people up. To do that cleanly, you need a few terms straight, because marketers throw them around interchangeably and they are not the same thing.

Start with the ICP, the Ideal Customer Profile. Your ICP describes the company, the account, that your product serves best. You build it from firmographics, which is just demographics for businesses: revenue, industry, company size, the tech stack they run. The ICP answers the question, what kind of organization should we go after? If you're business-to-consumer, or you're a solo operator selling to individuals, the analog is your ideal segment, but the logic is identical. You're describing the buyer at the level of a category, not a face.

Then there's the buyer persona. The persona is the actual human being inside that account, the person who decides or influences the purchase. A chief marketing officer. An IT director. You build the persona from psychographics: their goals, their pains, their professional fears, how they actually buy. So if the ICP is a mid-size software company with two hundred employees, the persona is the marketing director at that company who's terrified of missing her quarterly number.

Here's the crisp line to remember. The ICP is the what and the where. It's the company. The persona is the who and the why. It's the person. And they steer different decisions. Your ICP guides where you spend, which channels, which segments you chase first. Your persona guides what you say, the message, the angle, the words. Where to spend versus what to say. Keep those separate and a lot of muddy marketing gets clearer.

Now I want to add a sharper lens on top of personas, because personas alone can drift into stereotype. It's called Jobs-to-be-Done, JTBD, and it comes from the late Harvard professor Clayton Christensen, in his book Competing Against Luck. The core idea is that customers don't really buy products; they hire them to make progress on a job in their life. There's a canonical line that carries this, and it actually originates with Theodore Levitt, then Christensen carried it forward: people don't want a quarter-inch drill, they want a quarter-inch hole. Nobody wants the drill. They want the hole, and the shelf, and the tidy room, and the feeling of a done weekend.

Jobs-to-be-Done has three dimensions: functional, social, and emotional. The functional job is the practical task. The social job is how it makes you look to others. The emotional job is how it makes you feel. And the whole framework deliberately goes past demographics to get at why people actually switch from one thing to another.

Let me make that concrete with a contrast, because this is the difference between research that drives copy and research that just decorates a slide. A demographic persona says: thirty-five-year-old marketing manager. Okay. So what? A Jobs-to-be-Done statement says: when I inherit a content backlog with no strategy, I want to look in control fast, so I can keep my new boss's trust. Feel the difference? The second one practically writes the headline for you. It's got a trigger, a fear, and an outcome. The first one is just a row in a spreadsheet. The second drives copy. The first doesn't.

Next term, and this one is the heart of job one: voice-of-customer, VoC. Voice-of-customer is the exact words, the literal phrases, the emotional language that real customers use to describe their problem and the solution they want. And the goal here is simple but it takes discipline: you want to write your copy in the customer's words, not in the marketer's words. Marketers reach for words like streamline and optimize and unlock. Customers say, I was drowning, or, I just want my evenings back. Guess which one converts.

So how do you actually get voice-of-customer at scale? This is the central technique of job one, and honestly the spine of this whole research practice. It's called review mining, or message mining. Here's what it is. You go find the places where your target customers are already talking, unprompted, about the problem and about the solutions they've tried. And you harvest their literal language into a swipe file. Their words, copied down, kept verbatim.

Where do you mine? Lots of places. Software review sites like G2, Capterra, and Trustpilot. The Apple and Google app stores. Amazon reviews. Reddit threads, where people are brutally honest. YouTube comments. And then your own goldmines that you already own: your support tickets, your sales call transcripts, the open-ended answers from any survey you've run, and your win-loss notes from deals you've closed and lost. All of that is voice-of-customer waiting to be collected.

Now here's a pro tactic that separates people who've done this from people who've read about it. Mine the three-star reviews. Think about it. Your five-star reviews are euphoria, they just gush, this changed my life, amazing. Your one-star reviews are rage, garbage product, never again. Neither is that useful. But the three-star reviews? That's where the precise tradeoffs live. That's where someone says, I love that it does X, but I almost didn't buy because of Y, and I really wish it also did Z. The objections, the hesitations, the exact thing that nearly stopped the sale. Three-star reviews are the richest messaging material you'll find. Go there first.

And one more: don't only mine your own reviews. Mine your competitors' reviews, and reviews of adjacent tools in your space. Your competitor's three-star reviews are telling you what the whole market is frustrated by, and that's an opening you can position into.

Why does this matter so much commercially? Because practitioners report large conversion lifts from simply rewriting their pages in mined customer language. You'll hear claims floating around like seventy percent more qualified leads from this kind of rewrite. Treat that as a practitioner claim, a Copyhackers-style result, not a law of physics. Your mileage varies. But the direction is real and well documented: write in their words, convert better.

Okay, here's the load-bearing point of job one, and please tattoo this somewhere. AI is not the source of the language. AI is the organizer of language you collected. The correct workflow is: copy the real reviews, paste them in, and ask AI to cluster them. The wrong workflow, the one that wrecks people, is asking, hey AI, what do customers of product X say about it? That second question is hallucination bait. It will give you a beautiful, plausible, completely invented answer. Don't ask it to recall. Make it organize what you supply.

So let me give you the actual workflow, spoken so you can go do it. Workflow one, grounded review-mining synthesis. Your inputs are twenty to fifty real reviews, support tickets, or call transcripts, pasted in as raw text. You copy them by hand from G2, Amazon, Reddit, the app store, or you export your support tickets. Real text, no summaries.

Then, after you've pasted all those reviews, you give the model an instruction along these lines. You tell it: below are forty real customer reviews I copied from G2 and Reddit about this product category. Do not add any information that isn't in this text. Then you ask for four things. First, cluster the recurring pains into five to eight themes, and under each one, quote two or three of the verbatim phrases customers actually used. Second, list every objection or hesitation mentioned, in the literal wording. Third, pull the exact phrases people use to describe the outcome they want, the I just want to blank lines. And fourth, flag any theme that only appears once, so you don't over-weight a single loud voice. And you close by telling it: use only the customers' words, and if you're inferring anything, label it as an inference.

Why does that prompt work? Three reasons. It forbids invention outright. It demands verbatim quotes, so you get a swipe file, not a paraphrase that's lost all the juice. And it separates fact from inference, so you always know what came from a customer's mouth versus what the model guessed.

Now workflow two, building a grounded persona, or a Jobs-to-be-Done statement. Your inputs here are the clustered output from workflow one, plus any real data you happen to have: your CRM segments, your sales-call notes, your Google Analytics audience data, the real job titles of your actual customers. You feed all that in, and then you ask for one buyer persona, built only from that material. You require specific fields. The role or title. The trigger event that starts their search, the thing that happens that makes them go looking. Their top three pains in their own words. Their top two objections. The functional, emotional, and social job they're hiring you for, written in that Jobs-to-be-Done format: when this situation happens, I want to do this, so I can achieve this. And finally, where they go to get information.

And here's the crucial instruction, the one that makes this honest. You tell the model: for every field, cite which review or quote it came from. And if you don't have evidence for a field, write the words NO DATA, and do not guess. That NO DATA instruction is the antidote to generic personas. A persona that honestly says NO DATA on three of its fields is more useful, and more trustworthy, than a glossy invented one where every field is filled with confident nonsense. Empty boxes you can go fill. Fake boxes mislead you.

Let me name a few tools for job one, quickly, as interchangeable examples and not endorsements. There's SparkToro, an audience-research tool founded by Rand Fishkin and Casey Henry. It surfaces the podcasts, YouTube channels, subreddits, websites, hashtags, and search phrases that a defined audience actually engages with, using clickstream-type behavioral data rather than surveys, so it's what people do, not what they claim. In 2025 it added a natural-language Describe Your Audience feature, plus email-list upload and segmentation, and the free tier reportedly gives you around twenty searches a month. There's AnswerThePublic, which visualizes the questions and the prepositional phrases people search around a given keyword, great for seeing the shape of curiosity around a topic. And there's Perplexity, in its deep research or academic mode, which is an answer engine where every claim ships with a clickable source link, which makes it useful for market sizing and for finding real forums, precisely because you can verify each citation instead of trusting it.

That's job one. Gather real customer language, make AI cluster it, build a grounded persona with NO DATA where the evidence is missing. On to job two.

Competitor teardowns

Job two is the competitor teardown. The output here is a structured read on each competitor. You want their stated positioning, the category they claim. Their core value propositions and the proof they offer for them. Who they target. The content topics and formats they publish. The keywords they rank for. And then the highest-value part of the whole exercise: the gaps. The topics, the intents, the formats, the objections, the audience segments they completely ignore. Here's the line for that. Find what they're NOT saying. The silence is the goldmine, because that's the differentiated territory nobody's claimed.

Couple of terms first. SERP stands for Search Engine Results Page, just the page Google hands back for a query. And here's a subtlety: your SERP competitors, whoever actually ranks for your terms, may be totally different from your business competitors. A media site or a comparison blog might outrank your real rival for your money keyword. You have to know both.

Then there's search intent, which is the why behind a query. There are four classic flavors. Informational, where someone wants to learn, like how to optimize meta tags. Navigational, where they want a specific place, like a brand-name login. Commercial, where they're comparing before they buy, like best SEO tools 2025. And transactional, where they're ready to act, like buy SEO audit software. Intent governs everything downstream, because it dictates the format and the depth your content needs.

And then content gap analysis, which is finding the topics, keywords, questions, and formats your competitors cover or rank for that you don't. Here's the 2025 and 2026 framing, because this has evolved. Gaps used to mean just missing keywords. Now a gap is also missing entities, missing search intents, and missing formats. And increasingly, the most modern gap question is, are we even cited in the AI Overviews, are we showing up in the large language model answers? That's a semantic gap and an AI-visibility gap, and it's new territory you have to account for now.

So how do you run a teardown without the model confabulating? Same discipline as before. Ground it in real copy. Workflow three is the competitor messaging teardown. Your inputs are the actual copy you paste in. The competitor's homepage. Their pricing page. A few of their product pages. Their last ten blog titles. And their reviews from G2, Capterra, or Reddit. You copy the real text in. You do not, under any circumstances, ask the AI to recall the competitor from memory, because when you do that it confabulates features and invents pricing that doesn't exist. It will tell you a competitor has a feature they've never built. Paste the real text.

Then you instruct it: here's the real homepage and pricing copy from three competitors, each one labeled. Working only from this text, do four things. For each competitor, state their apparent market category, their primary value proposition, and who they target. Then build a comparison across them: category, top three claimed benefits, the proof or evidence they cite, and their tone. Then identify the themes all of them say, which are your table stakes, versus the themes only one of them says. And then, most important, list the pains, the objections, the audiences that none of them address. And you tell it: don't infer features that aren't in the text, and mark anything you're inferring.

Now there's a powerful second pass to this, and it uses the competitors' reviews instead of their marketing. Their marketing is what they want you to believe. Their reviews are the unfiltered truth. The G2, Capterra, and Reddit reviews expose the competitor's real weaknesses, the pricing complaints, the feature gaps, all the stuff you can position against. So you take those reviews and you ask: from these competitor reviews, list the top complaints and the unmet needs, in the customers' own words. Those complaints are angles your product could own. That's competitive intelligence you simply cannot get from looking at their website, because no company advertises its own soft spots.

Then there's workflow four, the content and keyword gap pass. Your inputs are either a keyword or SERP export from an SEO suite you already use, or a manual list of your competitors' article titles. And I have to be honest with you about where the line sits here. The cleanest gap data still comes from the dedicated SEO suites. Ahrefs, Semrush, and the rest have native Content Gap and Keyword Gap reports, and they're built on real crawl data. AI's job is not to invent that data. AI's job is to interpret the export. It's excellent at taking that list and scoring the opportunities, clustering them by intent, finding the patterns. It is terrible, and dangerous, at inventing keyword volumes.

So the prompt is: here's a list of keywords my top three competitors rank for that I don't, pasted in. Group them by search intent, informational, commercial, transactional, and by topic cluster. Flag the clusters that map to an actual buying decision. And for each high-value cluster, suggest the content format that fits the intent, and a one-line angle that differentiates from a generic article. And then the guardrail: do not invent search volumes, only use the numbers I gave you.

That guardrail matters because of the pitfall buried inside this job. AI will confidently make up keyword search volumes. It'll invent competitor traffic numbers. It'll tell you a competitor ranks number one for some term, all from memory, all fabricated. Any of those numbers must come from a real tool. The model can interpret numbers. It cannot source them. Hold that line.

That's job two. Paste real copy, find what they're not saying, and never let the model invent a volume.

Positioning angles

Job three is positioning. This is where research turns into a differentiated point of view: your value props, your angle, and the gut-check that kills generic claims. And there are three frameworks I want you to have in hand.

The first is April Dunford's positioning model, from her book Obviously Awesome: How to Nail Product Positioning So Customers Get It, Buy It, Love It. There's an updated edition out, too. Dunford breaks positioning into five components. One, competitive alternatives: what would the customer use if you didn't exist? And often the honest answer isn't a rival product, it's a spreadsheet, or it's nothing, they just live with the problem. Two, your differentiated or unique attributes: the capabilities only you actually have. Three, the value those attributes enable: the benefit the customer genuinely cares about, not the feature itself. Four, your best-fit customers: the people who care most about that particular value. And five, your market category: the frame of reference that makes your value obvious the moment someone hears it.

Here's Dunford's big thesis. Positioning is context. You deliberately choose the frame in which you are obviously the best choice for a specific set of customers. You're not lying, you're setting the stage so your strengths read as strengths. And the reframe she gives is gold for working with AI. Don't start from, what are our features. Start from, what would the customer use instead of us, and why are we better for them, specifically. That question reorients the whole exercise around the customer's real alternatives, which is where positioning actually lives.

The second framework is Eugene Schwartz's five stages of awareness, from his 1966 classic Breakthrough Advertising. Schwartz says you match your message to how much the prospect already knows. Stage one, unaware: they don't even know they have the problem. Stage two, problem-aware: they feel the pain, but they have no solution in mind. Stage three, solution-aware: they know solution types exist, but they haven't picked one. Stage four, product-aware: they know you specifically, but they're not yet convinced. And stage five, most-aware: they're ready, they just need the offer.

The consequence of this is enormous and most people miss it. The exact same product needs a completely different angle at each stage. A problem-aware buyer needs empathy and problem-naming; you have to show them you understand the pain before you ever mention your product. A most-aware buyer just needs the offer and the call to action; if you make them sit through your empathy paragraph, you lose them, they were ready to buy. So awareness is the bridge between your research and which angle you actually write. The research tells you the pains; the awareness stage tells you how much of the journey to walk before you make the pitch.

The third tool is the simplest and maybe the most useful day to day: the so what test. It's a differentiation gut-check. Take any claim you've written. Now swap your brand name out and drop a competitor's name in. If the sentence still reads as true, the claim is generic, and it fails. It's saying nothing. There's a sibling version called the name-swap test, same idea. So you keep pushing every value statement through, so what, so what, so what, until you reach a benefit the customer genuinely cares about and that a rival could not truthfully say about themselves. If we deliver enterprise-grade reliability, so what, every competitor says that, swap the name and it's still true, it fails. Push further until you've got something only you can claim.

So here's workflow five, generating positioning angles from grounded research. Your inputs are everything you've already built: the voice-of-customer pains and desires from job one, the competitor teardown and the what-none-of-them-say gaps from job two. You feed that in and you ask for an April-Dunford-style pass. One, what are the realistic competitive alternatives for this customer, including do nothing and the spreadsheet. Two, which of our attributes are genuinely differentiated against those specific alternatives. Three, translate each one into the value the customer actually named in the reviews, using their words. Four, propose three distinct positioning angles, each tagged with the awareness stage it fits, per Schwartz. And for each angle, write the one-line value prop, then run it through the so what and name-swap test, rewriting it until a competitor could not truthfully say the same sentence. And you tell it to reject any value prop that's true of every competitor.

The discipline here, and this is the whole game with positioning, is that AI suggests the angles but you judge them. You judge them against real customer language and a real competitive set. If you skip that judgment, you get positioning that's confident, plausible, and totally generic. The model is a brainstorm partner, not the arbiter of truth. You're the arbiter.

That's job three. Frame the context, match the awareness stage, and kill any claim that survives the name swap.

Content briefs

Now job four, and this is the keystone. The content brief is where every bit of research you've done lands and becomes operational. It's the artifact that makes a later draft dramatically better, and I'll lightly foreshadow here: in the later content-engine episodes, the brief is exactly what we'll feed into a drafting workflow. So learning to build a great brief now pays off twice.

Here's the mantra for this job, and really for the whole episode. Research that doesn't change the brief is research theater. Hold that. We'll come back to it hard in a minute.

Let me walk you through the anatomy of a content brief, the standard fields, because filling these out is the actual deliverable. There are eleven of them. One, your target or primary keyword, plus the secondary keywords. Two, the search intent, informational, navigational, commercial, or transactional, because that governs your format and your depth. Three, the audience, written as one sentence from your grounded persona, not a generic line, the real one with the trigger and the pain. Four, the angle, the unique point of view, the differentiated take from job three, the thing that makes this not a commodity article. And there's a real reason this field matters more now: the post-December-2025 Google core update reportedly weights Experience and the broader E-E-A-T signals harder, specifically to separate genuine human insight from commodity AI content. So your angle and your first-hand point of view aren't just nice, they're the ranking moat.

Continuing. Five, the outline, your required H2 and H3 structure with the key points under each. Six, competitive context, meaning the current top three results, their rough word counts, and what they cover, so you beat them instead of just echoing them. Seven, the sources to cite, real references, and this is where you bake verification right into the document. Eight, internal links, with the anchor text, the URL, and where it goes. Nine, your word-count or depth target, drawn from the SERP norms, what's actually ranking. Ten, the meta, your title tag at around sixty characters and your meta description at roughly a hundred fifty-five to a hundred sixty characters, with the primary keyword in there. And eleven, the voice-of-customer phrases to use, the exact customer language and the specific objections to address. That eleventh field is how you carry job one all the way into the draft. The customer's actual words ride in on the brief.

So workflow six is how you generate that brief from your research. Your inputs are the outputs of jobs one through three, the persona and Jobs-to-be-Done, the competitor gaps, your chosen angle, the voice-of-customer phrases, plus the real top three SERP results pasted in, plus your real internal-link list. And you instruct the model to turn all of that into a content brief for one article, filling every field. The primary and secondary keywords, using the keyword list you provided and not inventing volumes. The search intent. A one-sentence audience pulled from the persona. The differentiated angle from your positioning. An H2/H3 outline with bullet key-points under each heading. The specific objections to address and the exact voice-of-customer phrases to include. A competitive-context note based on the three real results you pasted. The internal links from your list with anchor text. A word-count target. And a draft title tag and meta description.

And then two finishing instructions that matter. You tell it: at the end, add a Sources to verify list. And, mark any claim that needs a citation with the tag VERIFY, in brackets. That VERIFY convention is doing real work. It carries forward, in writing, exactly which facts the writer has to check before anything gets published. It turns a vague feeling that I should fact-check this into a concrete checklist sitting right there in the brief.

There are tools that auto-generate briefs, and they're interchangeable: Frase, Surfer, Clearscope, SEOmonitor, plus the custom templates you can build inside your assistant. They genuinely speed up the structure, the skeleton of the brief. But, and this is the recurring caveat, the angle field and the voice-of-customer field still have to come from your grounded research. And anything those tools auto-suggest, the competitors say this, the people also ask that, should be spot-checked, because it can be wrong. Use them for speed on structure, not for truth on substance.

That's the four jobs. Now I have to spend real time on what goes wrong, because the failure modes here are specific and expensive.

Pitfall A, the big one: AI fabricates market data, competitor facts, and statistics

This is the one that ends careers, so let's be precise. AI will confidently invent statistics. It will cite studies that do not exist. It will attribute quotes to real, named people who never said them. It will get dates and places wrong. It will conjure competitor features and pricing out of nothing. And it will invent customer pains that no real customer has ever felt. And every bit of it arrives sounding authoritative.

Let me give you some numbers, and I'll tell you exactly how confident to be in each. One analysis found that AI-generated bad citations were around sixty-six percent total fabrications. Not corrupted versions of a real source, invented wholesale, from scratch. That's a Neil Patel hallucination study figure. And the fake-citation problem is reportedly growing, so mark these next ones as reportedly. In academic papers, roughly one in two thousand eight hundred twenty-eight papers had fabricated references in 2023. By 2025 that was about one in four hundred fifty-eight. By early 2026, about one in two hundred seventy-seven. And there's an estimate of around a hundred forty-six thousand nine hundred hallucinated citations across four major scientific repositories in 2025. Those come from a Lancet study reported by STAT News and phys.org. The trend line is the point: it's accelerating, not improving.

And it's not just academia. In marketing specifically, a reported forty-three percent or more of marketers say that hallucinated or false information has slipped past their review and gone public. Public. Out the door, on the website, in the campaign.

Here's the story to remember, the one that makes it real. In July 2025, a federal judge ordered two attorneys for the MyPillow CEO, Mike Lindell, to pay three thousand dollars each for an AI-assisted court filing that had more than twenty errors and cited cases that don't exist. And that was just one of more than two hundred documented court cases where AI-hallucinated citations got sanctioned. Sit with that. These are lawyers, trained professionals whose entire job is to cite things correctly, in front of a federal judge, with their license on the line. And the AI still sank them. The lesson writes itself: if it can sink a lawyer in court, it can absolutely sink a marketer's whitepaper. The stakes feel lower for us until the day a fabricated stat goes public under your brand.

Now, do the deep-research modes fix this? They help. They don't solve it. Let me explain both halves. The deep-research and agent modes, OpenAI's ChatGPT Deep Research, which runs multi-step web research and synthesizes hundreds of sources into a cited report; Google's Gemini Deep Research; and Perplexity Deep Research, which launched in February 2025 and runs dozens of searches in about two to four minutes, all of these attach citations to their claims. And that is genuinely valuable, because it lets you verify. Perplexity in particular is built around per-claim source links.

But here's the catch, and it's a sharp one. A citation existing is not the same as a citation being correct. The model can cite a real-looking URL that doesn't actually say what the model claims it says. It can mis-summarize a source that's real. It can surface some low-quality blog and present it as authoritative. So the honest framing is this: citations make verification possible. They do not make the output true. The link is an invitation to check, not a guarantee you don't have to.

So here are the rules to actually live by. Five of them.

Rule one. Click every link. If a stat or a competitor fact has no source attached, treat it as fiction until proven otherwise. And if it does have a source, open it, and confirm the source actually says that thing. Not that the link works. That the link supports the claim.

Rule two. Demand sources explicitly, in the prompt. Tell the model: cite a source for every factual claim, and if you can't, say unverified. You'd be amazed how much fabrication just declines to show up when you ask it to label its own confidence.

Rule three, and this is the single most important habit in this entire episode. Ground in real artifacts. Don't ask the model to recall the market, or the competitor, or the customer from memory. Paste in the real reviews. Paste in the real competitor copy. Paste in the real call transcripts and the real keyword export, and instruct it to use only that text. The phrase to remember is retrieval over recall. Retrieval, from documents you supply, over recall, from the model's fuzzy memory. Every workflow I've given you today is really just a version of this one rule.

Rule four. Separate fact from inference. Make the model label its inferences as inferences. And never, ever let a number originate inside the model. Numbers come from your sources. The model can move them around; it can't birth them.

Rule five. The two-source rule. For anything you're going to publish, confirm it in two independent sources. Not one link, two.

Those five rules are the immune system for this whole practice. Everything else is technique; these are survival.

Pitfall B: the generic persona, the one that describes no real human

The second pitfall is subtler because the output looks fine. Ask AI for a persona for some category with no real data behind it, and what you get back is an averaged internet stereotype. It's plausible. It's polished. And it's useless. The model has no access to your customer data, your behavioral logs, your context. So it hands you the statistical average of every blog post it ever read about that category, which describes no actual person.

And the danger isn't just that it's bland. It's that generative AI personas built without real data lead to confident, yet dangerously wrong, decisions. There's good writing on this, from Articos, from a UX Psychology piece, from Britopian, and from an ACM Interactions essay literally titled the Synthetic Persona Fallacy. The throughline across all of them is the same: treat AI personas without real data as hypotheses, never as facts. They're a starting guess to go validate, not a finding to go act on.

Let me give you the contrast so you can spot the difference instantly. The made-up persona reads like this: Marketing Mary, thirty-eight, busy, values efficiency, wants to grow her brand. Now, who is that? That's anyone. There's no quote in there. No trigger event. No objection. It's a collection of adjectives. The grounded persona, by contrast, has every field tracing back to a real review, ticket, or call quote. It includes the actual trigger event. It includes the exact objection language the customer used. And it carries a Jobs-to-be-Done statement written in the customer's own words. Plus, it honestly marks NO DATA wherever the evidence is missing.

Here's the tell, and it's quick. A grounded persona contains verbatim customer phrases, things a marketer would never naturally phrase that way. A generic persona contains adjectives the marketer would use, efficiency, busy, growth-minded. And the fastest way to recognize a generic one: it has no quotes, nothing in it surprises you, and it would fit your competitor's customers equally well. That last bit is just the name-swap test again, applied to a persona instead of a value prop. If this persona describes your rival's buyer just as well as yours, it describes nobody.

Pitfall C: research theater

And the third pitfall is the one that wastes the most time while feeling the most productive. Research theater. Here's how it goes. You run all the workflows. You produce an impressive persona deck and a gorgeous twelve-page competitor teardown. You feel incredibly productive, you've been busy for two days. And then the article you write reads exactly the way it would have read with zero research. Nothing changed. That's research theater: the performance of research without the effect of it.

So here's the standard. Research only counts if it changes a decision. The angle. The headline. The objections you choose to address. The actual words you use. If it didn't change one of those, it didn't happen, no matter how pretty the deck.

And there's a one-line test for it. If I deleted this research doc right now, would the draft come out different? If the answer is no, it was theater. Be ruthless with that question.

The fix is structural, not motivational. You don't fix research theater by trying harder. You fix it with the content brief, which is the forcing function. The research has to flow into specific brief fields: the angle, the voice-of-customer phrases, the objections to handle, the chosen awareness stage. And then the draft gets written from the brief. That's the chain. So here's the discipline: if a finding can't find a home in a brief field, it didn't matter, and you let it go. The brief is what converts research into decisions, and a decision is the only thing that changes a draft. No home in the brief, no impact on the page.

Where this is all heading in 2025 and 2026

Let me close with where the tooling is going, because two shifts matter for everything we covered.

First, deep-research agent modes are now standard across ChatGPT, Gemini, and Perplexity, as of 2025. Multi-step autonomous web research that produces cited reports. The top tier is tight; Gemini reportedly leads the DeepResearch Bench with the others within a few points, and you should mark those leaderboard specifics as reportedly, because they churn. The takeaway isn't which one wins this month. It's that cited, multi-step research is now table stakes, and your verification habits matter more than your tool choice.

Second, and this is the most relevant shift for our whole grounding thesis: connectors, and grounding in your own data. This is the productized version of paste in real artifacts. Claude Projects are persistent workspaces with an uploaded knowledge base. They accept PDFs, Word docs, CSVs, text files, HTML, and pasted text, with files up to around thirty megabytes, and the assistant references your documents in every single chat. Claude Connectors link it to Google Drive, Slack, and Notion. ChatGPT's Custom GPTs and Projects do the analogous thing with uploaded knowledge files and Actions. And Gemini Deep Research expanded in 2025 to pull from your Gmail, your Drive, and your Chat.

Think about what that means for everything we discussed. Instead of pasting forty reviews into every chat, you load your reviews, your transcripts, and your competitor copy once, and the assistant just stops guessing. It references your real material every time. That's the whole ground it in real artifacts idea, turned into a feature you configure once. And there's a stat that captures why this took off: reportedly, per a cited Gartner 2025 figure, around seventy-eight percent of businesses say generic chatbots fail at company-specific tasks. Of course they do. They don't have your data. These connectors are the fix.

And one more shift on the search side. AI Overviews and large language model answers mean your content gaps now include a brand-new question: are we even cited by the AI? And Google's December 2025 core update reportedly leans harder on first-hand Experience, that E-E-A-T signal, which reinforces the thing we keep coming back to: your angle and your real point of view are the moat now, not sheer volume of content. Commodity content is losing. Grounded, opinionated, first-hand content is winning.

So here's your homework, and I want you to actually do it, because this stuff only sticks when you run it on your own brand. Pull thirty real reviews of your product, or your closest competitor if you're new, and run workflow one. Count how many pains surface that you didn't know about. Then paste your top competitor's homepage and pricing into workflow three, and find one thing none of your rivals say. Then turn all of it into one brief with workflow six, write the article from that brief, and compare it to your last from-scratch article. And finally, run every persona field and every value prop through the name-swap and so what test, and delete anything a competitor could also truthfully say.

That's the loop. Gather real material, make AI structure it, verify everything, and land it in a brief that actually changes the draft. AI is fast at structuring and synthesizing. It has no idea what your customers said. So your job, the whole job, is to bring the real words to the table and refuse to let the model make them up. Do that, and you've got a research engine. Skip it, and you've got confident fiction. Go build the engine.