ArticleIssue #174 | What To Actually Do About AI Search
This week’s issue came out of a client conversation I had on Monday – I was talking to an incredibly smart, well-read CMO who said (paraphrasing): I’ve read thousands of pages on AI stuff. I know it’s important. I hear from our customers that they’re finding us via ChatGPT. And I have NO IDEA what concrete steps to take to help us show up more. I have lots of theories and articles and tin-foil-hat conspiracies, but no checklist I can give to our team.
That got me thinking – didn’t I write about this? It turns out, way back in February, I wrote The Click Isn’t Dead. The central idea was simple: AI didn’t kill the click; it relocated it. That issue got more replies than any other issue in 2026 (at least, to date). It also resulted in ~17 responses, all asking some variant of the same follow-up question: “Great. So what do I actually do about it?”
The question is both fair AND difficult to answer, because the honest answer is simultaneously: (1) we’re still not entirely sure and (2) what we are relatively confident about is subject to continual change. Showing up in AI search is less a “thing you do” and more like trying to build a house of cards on a speedboat – just when you think you’ve got it, you hit a wave and the whole thing comes crumbling down.
That being said, there are some things I’ve noticed since writing that original article: the AI search conversation has produced an astonishing volume of philosophy and an almost total absence of homework. Everybody’s writing a LinkedIn thoughtboi article or a bullshit framework – and almost nobody is writing the to-do list. My bookmarks tab is full of think-pieces on vector embeddings and passage-level retrieval and “agentic friction” …. but none of that can be sent to a client or brand owner who is struggling to figure out how to get their content/site to show up when a relevant prospect is in need of a solution they provide.
The idea of this issue is to begin to address that gap, in simple, easy-to-understand and easier-to-execute terms. No acronyms to memorize, no crazy theories, no technical SEO gibberish – just the work we’ve found to be successful in getting both ourselves and our clients to show up in LLMs.
My one disclaimer before we dive in: most of it is profoundly unglamorous. It’s boring. It’s tedious. It’s – to some degree – basic. But 80% of marketing success is being brilliant at the basics, so it makes sense that the same applies here.
Let’s get to it.
Step #1: Ask the Machines What They Say About You
Start here, because almost nobody has, and it’s free.
Open ChatGPT, Gemini, Claude and Google’s AI Mode, and run the 15-20 questions a real buyer in your category would actually ask. Not “is [your brand] any good” – nobody actually types that. Focus on the questions real people are likely to ask: “best [product] for [specific situation],” “alternatives to [your biggest competitor],” “is [your brand] worth the cost?” “Best solution for [common pain point/challenge]” – the questions a real person asks when they’re close to spending money but not quite sure yet.
If you’re not sure, then do some actual audience research – go ask your friends/family what they’d search. Call (yes, with a phone) 5-10 of your best customers and ask them how they’d go about asking ChatGPT about a specific problem you solve for customers/clients. Then, head over to SparkToro and use their AI search intelligence tool to validate what you’ve found or identify the gaps. The combination of 1P data (what you asked people) and 3P data (the observed stuff from SparkToro) is what provides robustness + coverage.
Once you have your questions, run them all through each model (also, you can use SparkToro + audience questions (seriously, if someone says they found you via AI, train your call center/sales team to immediately follow up with, “which AI tool did you use?”) to figure out which one(s) your audience uses most, then use that to weigh responses from each. The key is to do this MULTIPLE TIMES at different times of the day, from different devices in different statuses (logged in, logged out; premium vs. free). The single-most-common mistake I find is brands running a search on a non-deterministic model once, then treating the output as canonical. It isn’t. You must run it multiple times to get a coherent picture of the landscape. Put all the responses into a Google Sheet, dated for the time period when you ran the searches.
Armed with that intel, read each response the way a customer would. Focus on 5 questions:
- Do you show up at all?
- If yes, what content about your brand shows up? Is that content accurate?
- Which of your competitors also show up / show up in place of you?
- What does it say about each of those competitors? Is that content accurate?
- What are the most common sources of information (reddit, linkedin, news articles, industry blogs, etc.)?
That 3rd question provides your real competitive set in the AI ecosystem. It will vary from platform to platform – but (increasingly) you’ll find that it is both more accurate to who you’re actually competing against AND less aligned to who you think you’re competing against. The reality is that more people are using AI to inform more decisions every day – so the other firms/groups AI *says* are your competition increasingly shapes who you compete against.
My advice is to do this yourself the first time. I have subscribed to or trial-ed a half dozen “AI Visibility” tools – and nothing produced by any of them was as illuminating as the 2 hours I spent one Thursday afternoon doing this myself. It sounds crazy (and I feel crazy writing it), but it’s the truth: forcing myself to do this was critical.
AI Visibility Tools are – mostly – garbage. They cherry pick prompts, they often take a snapshot of the responses vs. running the same query multiple times, they tend to favor the kinds of queries that make the platform *appear* useful vs. actually helping you improve visibility.
Just do it yourself to start. Then build an agent to do it for you. You’ll save money and get better data.
Step #2: Identify The Relevant Source Material
The #1 reason I ask that 5th question is simple: when an LLM answers a question about your category, it (usually) isn’t conjuring an opinion out of the ether. It’s assembling that content from actual sources – videos, pages, podcasts, blogs, local listings, reviews, Reddit threads, something. When I did this the first time, almost half the errors were generated based on outdated content from our own site (they were ~10 years old, but still!). If the answer about you is wrong, thin or generic, the cause is usually sitting right there on your site or on another property (like your subreddit or company LinkedIn) you can control.
That means the highest-leverage thing you can do isn’t technical at all – it’s to go read your own pages the way a model ingests them (flatly, literally, looking for the answer to a specific question), then edit them accordingly. Make it easy for an LLM to ingest the content and get the right answer. If you’re not sure, take the URL and put it into the model that got the information wrong, and ask it why/how to fix it.
The solutions are almost embarrassingly basic:
- Answer the actual question(s) early and plainly, because models extract clean, direct statements; if your PDP buries “how much is it” and “who is this for” under 3 paragraphs of fluffy language your brand team loves, the only thing you’ve accomplished is making it more difficult for AI to understand (because let’s be real: your customers/clients aren’t reading it either).
- Add answers to questions you don’t have – if your research revealed that customers/clients are asking questions that your site/content doesn’t answer, create content that does. We’ve done this research hundreds of times for dozens of brands, and every single time, the questions the marketing/brand team *thought* were relevant and the questions their actual customers/prospects *told us* were relevant differed materially. If your site doesn’t answer the questions people are asking, you’re not getting cited.
- Serve the facts in text – specs, pricing, use cases,comparisons, FAQs – not in an image or a PDF or a video with no transcript. Parsing unstructured data (like video or audio clips or images) is compute-intensive (read: expensive). LLMs aren’t going to spend more just to help you.
- Eliminate your own contradictions: the name of the game is pattern recognition. Everything you say either supports the pattern you’re trying to build or detracts from it. If your homepage, your PDP and your reviews tell a slightly different story, you’re (inadvertently) making the LLM’s pattern of your brand less clear. And the less clear the pattern, the more likely it is that it will reach for someone clearer in answers (this is the same thing you probably do – if a friend asks for a referral, and you know of 2 people – one who you are SURE does it, and another that you kinda think does it, which name are you going to give? The “sure” one – at least, provided you like the person asking for the referral).
This is 80% of the work…it just so happens to be the least sexy work imaginable. How you show up in an AI answer is a product of what’s on your site long before any “optimization” enters the picture – which, if you’ve read me for a while, is just the AEO version of a drum I’ve been banging for years: most of what happens in the machine is a product of what happens outside the machine.
Step #3: Become A Source, Not A Summary
If I could get you to internalize one thing from this whole issue, it’s this: LLMs cite originality.
That isn’t conjecture. It is a core tenet of the patents that Google, Microsoft, Perplexity and Anthropic have filed. The best description of it is “Information Gain” (filed by Google in multiple patents) – the idea that LLMs reward sources that provide net-new information to the model over ones that simply re-state information that could be obtained from multiple other sources.
And logically, that makes sense! If 100 sites all have the same information, from the LLM’s perspective, 2 things are true: (1) that information has a higher probability of being accurate and (2) there’s no need to cite all 100 cites – just cite the most defensible, most credible of the bunch and ignore the other 99. The harsh truth is that, for almost everyone reading this, your site is going to be in the ignored 99 UNLESS you add something to the view: original data, a real point of view, a proprietary methodology, a counter-intuitive perspective, real, demonstrable expertise – something the LLM can’t get anywhere else. That’s information gain.
The practical implications of that are massive: if your content is a competent rewrite of the current top 10 results (and most content on most sites is) then odds are you’re going to find yourself in the 99 ignore results way too often. That means your brand is dissolved into the answer with no attribution, because you didn’t add anything the model couldn’t get 100 other places. The brands that get named are the ones publishing things that didn’t already exist on the internet: a benchmark study, original research, a survey of their own 200 customers, a genuine “we looked at X and here’s the crazy thing we found,” a clear, unpopular/counterintuitive position in a sea full of people hedging.
So before you publish anything, ask it the only question that matters: does this add something to the internet that an LLM can’t get anywhere else?
If no, you’ve spent a ton of time and resources creating something that – if you’re lucky – will be indexed then ignored.
If yes, you’ve handed the LLM a reason to cite you specifically. When the cost of producing competent content falls to zero (and it is), originality becomes immeasurably valuable.
Step #4: Go Deep on a Few Things, Not Wide on Everything
The instinct in a moment like this is to panic and produce – more articles, more “optimized” content, more stuff – essentially, take the Meta Ads approach to LLMs (i.e. hit publish a 1,000 times and hope volume saves you). Save yourself the trouble and suppress that instinct. Volume of mediocre content is free for everyone, and a thing that’s free for everyone is worth nothing to anyone.
The move is the exact opposite. Pick a small number of themes you can own more completely than anyone else in your category, and go so deep on them that you become the obvious answer. A model learns to associate you with a topic the same way a person does – through repeated, substantive, consistent presence, not through a scattershot of shallow pages touching everything once. Being authoritative on 3 things > Being “present” on 30.
If you want a concrete example, look no further than Rand Fishkin. He *could be* an authority on any of 100 topics from SEO to user behavior to audience research – but he’s consolidated everything he does around a single core theme: “discovery is consolidating into a handful of zero-click walled gardens, attribution and rank-tracking are collapsing as measurement constructs, and the only durable response is brand and business-model defensibility rather than tactical optimization.” The overwhelming majority of the content he produces ladders into that conclusion in some way, shape or form. That is what depth looks like – everything connects up to a few core ideas.
Step #5: Make Your Expertise Obvious to People First & Machines Second
There’s a real, do-it-this-week version of all the “structured data” advice flying around the interwebs that doesn’t require you to hire an SEO firm or learn what a vector is. It’s actually pretty simple:
- Put real names and real credentials on your content – An actual expert with an actual bio beats an anonymous “Team” byline, for human trust and machine citation alike.
- Write your headings the way people actually ask things (“How much does X cost?” beats a clever pun every time, because it matches the question being retrieved).
- Add the basic structured data – organization, product, FAQ, author markup. This is not some magical panacea – it’s table stakes. But it needs to be done, so do it.
- Add in a llms.txt file to your site – this is NOT a replacement for robots.txt (which tells crawlers where they can’t go); it’s an emerging, open-source standard that helps AI agents find your most valuable content.
- Ensure your content has simple, easy-to-extract answers for the basic/common questions – the mistake most brands make is embedding the answer within a larger paragraph (which forces the machine to extract/infer the answer – that introduces both error and effort, neither of which is good). Take the old Geico approach: make answers so easy a caveman can find them.
- Clearly and obviously state your differentiators over and over again. That’s easier said than done – one of my biggest challenges as a writer is my aversion to saying the same thing twice. I remember everything I’ve ever written – and for a long time, I tried to avoid saying the same things over and over again. It took Aaron Orendorf literally yelling at me to change that. The easiest test: if writing something one more time makes you want to throw up, you’re just starting to say it enough.
As you do this, remember – you’re not trying to trick a machine. You’re trying to make genuinely expert content look as expert as it is. People reward exactly the same thing. Solve for the people and the machines will follow; solve for the machines, and the people will ignore you.
Step #6: Build On The Platforms Your Audience Trusts
As you go through your Step #1 data, you’re likely to find that a disproportionate number of responses come from a select few 3rd party platforms – Reddit, LinkedIn, Wikipedia, Quora, X, YouTube, etc. That list is a map of where the model sources its credible information in your space. It (almost always) will differ from the platforms your brand team assumes matter.
The honest reality: you have to show up on the 2-3 that dominate your results, because that’s where your audience is asking and where the model is getting its information. But – and this is the part most people skip – showing up doesn’t mean posting more. It means running the exact same playbook from Steps 3 through 5 on someone else’s real estate.
A Reddit answer that genuinely, completely solves the problem in the thread (not a drive-by link drop) is information gain on a surface the model already trusts. A LinkedIn post with a real, specific point of view beats 10 that restate consensus or state nothing interesting.
Getting your company into the reference-grade sources that feed Wikipedia entity data does more than a year of blog posts. A YouTube video or Podcast is worthless to a model without a correct, easy-to-cite transcript, and priceless with one. This is the same game you’ve been playing on your own site, just run on a different field. It isn’t an either/or thing, it’s a both-and thing – do it on your site + the sources you control AND do it on the relevant 3P websites.
Step #7: But Don’t Hand Your Audience To The Landlords
One warning before you go execute all of this: every hour you spend optimizing for an AI surface (whether that’s AIO in search or a ChatGPT response) is an hour spent on – essentially – tenant improvements. You’re investing your time & treasure improving real estate you don’t own and can’t control, on a platform that will rewrite the rules on a whim with no notice. None of that is to say that it isn’t necessary and worthwhile work – but if it’s all you do, you’re building a castle on rented land. And, as anyone who built a business on organic Google pre-Panda will tell you, that movie doesn’t end well.
The rational response isn’t to avoid building on rented land; it’s to pair every strategy above with one that drags those hard-won visitors into something that’s actually yours: an email list, a community, a newsletter, a podcast subscription, anything – so long as it is a reason to come back to you instead of back to the answer engine time and time again.
Your initial goal should be to get cited (it is absolutely critical) – but you must have a strategy for converting that initial point of leverage into a relationship you control – one that no platform can revoke on your behalf.
The Takeaway
Strip away every acronym and the AI search playbook turns out to be almost insultingly familiar: be a real source, say true things clearly, go deep on what you can own, make your expertise obvious and never stop pulling people into relationships you control. The thing reading your content changed. What makes content worth reading is the same as it has always been.
So this week, do exactly 1 thing: go ask the machines what they say about you. Everything else on this list starts with what you find when you do.
One quick housekeeping note: I’ll be traveling for the 4th, so next week’s issue is going to look a little different. Instead of the usual deep-dive, I’m putting together a roundup of the 5 most popular/most cited issues. Consider it a long-weekend reading list, and a chance for me to point back at the ideas that clearly resonated most.
And while I’m at it: thank you. Whether you’ve been here since Issue #1 or you signed up last month, the fact that nearly 10,000 of you keep opening these, commenting & sending me your thoughts is the entire reason I keep writing them. It’s one of the best parts of the week. I don’t say it often enough, so: thank you for reading.
Cheers,
Sam
