Last week I made the case that “AI-friendly content” is mostly just better content: clear structure, real answers, honest sourcing, nothing exotic. That still holds. But it leaves a bigger question unanswered, one I hear from almost every internal client and corporate marketing colleague lately: “How can we get AI to recommend us?”
I struggle with that a bit, because I usually think it’s the wrong question. Not because AI visibility doesn’t matter, but because the desire to “impress AI” treats these tools like judges you’re performing for. They’re not judges. They’re a fast, oddly candid research panel that will tell you, in plain language, what it currently believes about you, if you just ask. The useful move isn’t to court that panel. It’s to interview it, write down what it says, and use the answers to fix what’s actually wrong.
Know which fight you’re actually in before you start swinging. A model that already knows you and ranked you low is a very different problem than one that’s never heard of you at all.
What these tools actually are, for this purpose
ChatGPT, Google’s AI Overviews and Gemini, Perplexity, Copilot… all of them are stitching answers together from your own content, whatever other people and publications have said about you, and whatever they can pull in real time. Ask one a question and what comes back is a snapshot of your reputation as a machine currently sees it. A little unsettling, honestly, and also some of the most useful free research you’ll do all quarter, if you know how to read it for what it’s actually worth.
And they don’t all read the room the same way. That’s not a footnote, it changes how you should run the test I’m about to walk through, and it’s more interesting than it sounds. More on that in a minute.
Three things worth checking, not one
Most people stop at “am I mentioned.” That’s not enough.
- Comprehension: does the model understand who you are, what you do, and who it’s for, when you ask directly?
- Positioning: when someone asks the broader question, about the problem you solve or the category you’re in, without naming you, do you show up at all? A featured answer, or a footnote?
- Sentiment and accuracy: when you are mentioned, is the description accurate, current, and clearly you, not confused with someone who shares your name or industry?
Answer those three honestly and you’ve basically answered the only question that matters here: does what AI currently says about you match reality, and does it match what a real customer needs to hear before they pick you over the next name on the list?
A test you can run yourself today
Here’s the process I actually use, stripped of anything specific to any one client, because the value was never in some secret method. It’s in actually doing this, on purpose, in writing, instead of reacting to whatever chatbot screenshot someone dropped in Slack this week.
Pick five or six real questions a prospective customer might actually ask. Not about you by name. About their problem. I run five discovery scenarios, plus one I’ve learned to always tack on afterward:
| Scenario | Prompt Shape | What It Tests |
|---|---|---|
| Named-brand | “What does [brand] do, who is it for, and what problems does it solve?” | Whether the model understands you when you’re named directly. |
| Category | “What are the best [your category] for [core buyer problem]?” | Whether you show up unprompted, without being named. |
| Pain-point | “How do I solve [specific problem], without naming any vendor.” | Whether you’re associated with the outcome, not just the label. |
| Competitive | “Compare [brand] to [two or three real alternatives].” | How you’re framed once you’re already part of the conversation. |
| Vertical / niche | The pain-point question again, aimed at your best-fit audience segment. | Whether your ideal customer specifically finds you. |
| Follow-up (always do this one) | “Why didn’t you mention [brand]?” or “What about [brand]? Have you heard of them?” | Whether the model knows you and just didn’t rank you, or genuinely has no signal on you at all. |
That last row does more work than it looks like it should, and it might be my favorite part of the whole exercise. Name yourself, and if the model backpedals with “Yes, actually, [brand] does…”, it knew you and simply didn’t reach for you unprompted. That’s a ranking and framing problem, and it’s usually the faster of the two to fix with content. If it still comes up empty, or fumbles basic facts even after you name yourself, that’s an entity problem: nothing it’s drawing from has a clear, consistent picture of you anywhere. That one takes longer, but it’s better to know which fight you’re actually in before you start swinging.
Run each prompt, exactly as written, across ChatGPT, Google’s AI Overview or Gemini, Perplexity, and Copilot. Same prompts, same order, every platform. Score four things for each answer: did you show up at all; how prominently, a featured pick or one name buried in a list of twelve; was the description actually accurate; and who got named instead of you, or alongside you.
Then ask one more question on every platform, especially Perplexity and Copilot: “What are your sources for that?” You’ll sometimes get back a page you forgot existed, an old press release, a stale directory listing, a partner’s case study mentioning you in passing. Genuinely useful, unprompted intel on who’s currently speaking for you online, whether you asked them to or not.
Where the actual prompts should come from
The five scenarios above are a shape, not a script. The part that actually matters is what you plug into the brackets, and that shouldn’t be generic.
Think about where you already suspect a soft spot: a vertical you want more of, a use case you’re genuinely strongest at but never get credit for, a competitor set you keep losing pitches to. That’s what belongs in your category and pain-point prompts, not just your product label swapped in for [brand].
This is exactly the kind of thing that showed up in a real audit I ran recently (details scrubbed, for obvious reasons): the client had a specific industry vertical they badly wanted to grow into, and that vertical turned out to be the cleanest miss in the entire test. Multiple AI platforms built out detailed, competitor-heavy shortlists for that exact use case, and the client’s name showed up in none of them.
Then ask “what are your sources for that?” on every one of those answers, especially on Perplexity and Copilot. It’s the fastest way to catch the sentiment and accuracy problems from the three checks earlier: a stale claim, an outdated number, or an answer that’s quietly describing someone else who happens to share your name.
Here’s the part worth sitting with, though: knowing that didn’t fix anything by itself. The test just pointed a finger at where the gap was. Somebody still had to go write the page, make the case, and give the model something to actually find. A prompt test doesn’t create the content any more than a blood test cures the thing it flags. It just tells you where to point the shovel.
Why the platforms don’t agree, and how to use that
Here’s the part nobody explains well, and it’s worth five minutes, because it changes where you go looking when an answer disappoints you. This is roughly how each of these systems works as I’m writing this, in mid-2026; the exact mechanics will keep shifting as these companies ship updates, so treat it as a map, not something to tattoo on your arm.
| Platform | What it’s actually grounding on | What that means for you |
|---|---|---|
| ChatGPT (search on) | Bing’s web index for general queries, plus direct publisher deals for news, sports, and finance. | If you’re invisible here, check whether Bing has indexed and understood your site before assuming it’s a Google problem. |
| Copilot (consumer) | Also Bing’s index, retrieved and reasoned over through Microsoft’s own orchestration layer. | Shares ChatGPT’s Bing dependency. A Bing fix often shows up in both places at once. |
| Copilot (Microsoft 365, enterprise) | The organization’s own Microsoft Graph data first, SharePoint, Teams, email. Open web is a separate, admin-controlled toggle. | Matters most if your buyer is doing research from inside a company that runs on Microsoft 365. |
| Google AI Overviews / Gemini | Google’s own search index. | The closest thing to “normal SEO” still being the whole job. Crawlability, indexation, and clear titles apply directly. |
| Perplexity | Its own independent retrieval and ranking pipeline, not Google’s or Bing’s. | Can find and cite you before Google or Bing catch up, and can just as easily miss you when you rank well everywhere else. |
OpenAI’s own announcement says ChatGPT search “leverages third-party search providers, as well as content provided directly by our partners”, cagey by design, but it’s been consistently and independently reported, including by CBS News, that Microsoft’s Bing is that provider. Consumer Copilot runs on the same rail. So if you’re invisible on both, don’t start the investigation in Google Search Console. Start in Bing Webmaster Tools instead, since there’s a real chance nobody there has properly indexed you yet.
Enterprise Microsoft 365 Copilot is a different animal. Microsoft’s own documentation is explicit that it grounds first in your organization’s Microsoft Graph data, the SharePoint, Teams, and email your buyer’s company already has, before web search even enters the picture as a separate, admin-controlled toggle. That matters if the person researching you is doing it from inside a company that runs on Microsoft 365, which, statistically, a lot of your buyers probably are.
Google’s AI Overviews and Gemini ground in Google’s own search index, full stop, and Google says as much directly. This is the one place where “normal SEO” and “AI visibility” are close enough to the same job that you can stop overthinking it.
Perplexity runs its own retrieval and ranking pipeline, independent of Google or Bing, and its own help center walks through roughly how. That independence cuts both ways. It can find and cite you before Google or Bing catch up, and it can just as easily miss you when you’re already ranking well everywhere else.
A gap on one platform and a strength on another isn’t a contradiction. It’s a clue about which index actually has a clear picture of you, and which one doesn’t.
Two examples, so this isn’t abstract
I’m going to invent two businesses to run through this, because staring at another table full of [brand] in brackets is nobody’s idea of a good time. Neither of these is real, though I tried to make them sound like the kind of businesses you’d actually run into in a real search.
Jones Roofing Specialists is a local roofing and solar company. Crestline is a national B2B project management platform built for construction general contractors. Same five-scenario framework, aimed at two very different buyers.
| Scenario | Jones Roofing Specialists (local, B2C) | Crestline (national, B2B) |
|---|---|---|
| Named-brand | “What is Jones Roofing Specialists, and what do they do?” | “What does Crestline do, who is it for, and what problem does it solve?” |
| Category | “Who are the best roofing companies near [city]?” | “What’s the best project management software for construction general contractors?” |
| Pain-point | “How do I know if I need a roof repair or a full replacement after storm damage?” | “How can a construction company reduce change-order disputes and keep subcontractors on schedule?” |
| Competitive | “Compare Jones Roofing Specialists to [Competitor A] and [Competitor B].” | “Compare Crestline to [two named competitors].” |
| Follow-up | “Why didn’t you mention Jones Roofing Specialists?” | “Why didn’t you mention Crestline?” |
Jones Roofing Specialists (local, B2C)
Run a set like this for a local service business and the shape is almost backwards from what you’d expect. The category prompt often does better than the named-brand one, especially on Google’s AI Overviews and Gemini, because local intent leans hard on reviews, map listings, and directory data the model can grab directly. The real miss shows up on the pain-point prompt: the model will happily explain how to tell a repair from a full replacement, in general terms, and never once put a name to it, because nobody wrote that explanation and signed their name to it.
Crestline (national, B2B)
This is the pattern I keep coming back to, and it’s the one I see most often with established B2B companies: strong, accurate answers on the named-brand and competitive prompts, because somewhere there’s a clear, well-linked page saying exactly who Crestline is and how it stacks up. Then near silence on the category and pain-point prompts, because nothing on the site is written to answer “what’s the best tool for this” without already knowing to ask about Crestline by name.
What the pattern usually looks like, at scale
Run this for almost any established business, real or make-believe, and that same shape shows up again and again. Named-brand comprehension is strong; ask an AI system directly who you are and it usually nails it, sometimes down to a stat or claim it lifted straight off your own site. Category and pain-point questions are where it falls apart. The same business that gets described accurately when named quietly vanishes the second the question is about the problem instead of the brand. Competitive comparisons, when a competitor set is already spelled out in the prompt, tend to be the strongest results in the whole exercise, because the model’s been handed the frame and only has to slot you into it correctly.
That gap, understood when named, invisible when not, is the real finding almost every time. It isn’t a technical failure. It’s evidence that your content answers “who are you” a lot better than it answers “why should I care, and why you instead of them.”
Turning evidence into action
This is where “impressing AI” and “using AI as evidence” actually split. One path rewrites content to please a black box: stuffing in FAQ schema, sprinkling in “AI-optimized” language, crossing your fingers. The other path reads the test results like a diagnostic and goes and fixes the specific thing they’re pointing at:
- Identity right, category missing: your comparison and category content likely has an answer gap, not a technical one. Build pages that plainly answer the buyer’s problem and name your category, not just pages that name you.
- It knew you once you named yourself: that’s a ranking and framing problem, not an entity problem. Clearer category and pain-point content is usually the fastest fix on this list.
- Competitor set surfaces cleanly, you’re absent from it: check whether you actually have a page making that exact comparison. Models tend to reuse whatever framing already exists on the web; if nobody’s written the comparison, the model has to build one without you in it.
- Mentioned, but described inaccurately or with stale numbers: that’s usually your own site’s fault, contradictory claims or inconsistent positioning across pages the model is stitching together.
- A vertical or niche test comes up empty: that’s often the clearest, cheapest opportunity in the whole exercise. A dedicated page for that specific audience, in their language, tends to close the gap faster than anything else on the list.
None of that is an AI trick, by the way. It’s the same content and positioning discipline that’s always mattered, just aimed with better evidence at exactly where it’s currently failing.
One caution
Treat any single run of this as a snapshot, not a verdict. Answers shift with model updates, retrieval behavior, even the day you happen to ask, and the sourcing details above will drift too, these platforms don’t sit still for long. Run it once to find your real gaps. Run it again on a quarterly cadence to see whether the work you did actually closed them. One odd answer from one chatbot on one afternoon isn’t evidence of anything. A consistent pattern across four platforms and multiple prompts, repeated a few months later, is.
The actual reframe
You don’t need AI to like you.
You need to know, honestly, what it already believes, and whether that lines up with the truth and with what your actual customers are trying to find. Once you know that, the fixes are almost always the fixes that were always worth making: clearer positioning, honest comparisons, content that answers real questions instead of performing for search engines or chatbots. AI just makes the gap visible faster than it used to.
None of this is about getting more out of AI, faster. It’s about being honest with yourself about what these tools already believe, and putting your actual time, and your content and positioning effort, where the evidence says it needs to go.
If you run this yourself and something turns up you’d like a second, experienced set of eyes on, reach out. Take a look at what’s on the table at ericlander.com/services in the meantime.