How Does ChatGPT Source?

Your brand might already be invisible inside ChatGPT, and you would not even know it. Every source it pulls from lands in a hidden trust tier, and most companies never make the cut.

This episode breaks down how ChatGPT decides what to fetch, what to cite, and what to ignore, plus the simple fixes that move your pricing and product facts into the tier that actually gets trusted. Where does your brand rank today?

Transcript

SPEAKER_01 0:13

Um usually when a business pours a massive budget into a beautiful, you know, high-end website, there's this fundamental assumption that your primary customer is actually going to be able to see it.

SPEAKER_00 0:23

Aaron Ross Powell Right. Yeah. You build it for human eyes.

SPEAKER_01 0:26

Aaron Ross Powell Exactly. You want stunning graphics, maybe an an interactive pricing slider, complex drop-down menus. I mean, the works.

SPEAKER_00 0:34

Aaron Powell And designing an experience for a person has basically been the gold standard since the internet began. We we optimize for the human click.

SPEAKER_01 0:42

But the moment you step into the world of AI search engines and uh autonomous agents, that entire assumption just completely shatters.

SPEAKER_00 0:50

Aaron Powell Oh, absolutely. It falls apart.

SPEAKER_01 0:51

Aaron Ross Powell Because we are operating in a digital landscape right now where your most important, most lucrative buyer might just be a machine.

SPEAKER_00 0:58

Yeah.

SPEAKER_01 0:59

And that machine might be entirely blind to the masterpiece that your agency just built. So we're taking you, the marketing leader, the agency owner, the business strategist, into the completely hidden machinery of AI search engines.

SPEAKER_00 1:12

Trevor Burrus, it's a totally different world under the hood.

SPEAKER_01 1:14

It really is. Yeah. And our mission for this deep dive is to answer a really urgent question. When AI looks for answers about your business category, what source pipeline is it actually using? And honestly, uh, does your brand even stand a chance of being pulled in?

SPEAKER_00 1:30

And to map this out, we're looking at two highly revealing analyses today. Right. The first is a study that deployed a simulated AI agent to navigate the websites of a hundred top B2B software products.

SPEAKER_01 1:42

Just to see what it would do.

SPEAKER_00 1:43

Exactly. The researchers wanted to see what would happen if an AI tried to perform a basic buyer task. And then the second piece of the puzzle is this under-the-hood breakdown of Chat GPT's raw network traffic.

SPEAKER_01 1:55

Which is wild to look at.

SPEAKER_00 1:56

It is. It exposes the secret internal labels the AI uses to choose its citations before it ever types out a single word to the user.

SPEAKER_01 2:04

Aaron Powell I want to start on the front lines with that B2B software study because the results were, I mean, an absolute bloodbath.

SPEAKER_00 2:10

Aaron Powell Yeah, that's a good word for it.

SPEAKER_01 2:12

The researchers gave an AI agent a very simple, high-intent buyer objective. Go find the monthly pricing and the core features for these 100 major software products.

SPEAKER_00 2:22

Aaron Powell Which is exactly the kind of task a procurement officer or or a potential client would delegate to an AI today.

SPEAKER_01 2:30

Trevor Burrus Right. Like compare these three platforms and tell me what they cost. Simple, right?

SPEAKER_00 2:34

Aaron Powell You'd think so. But the agent hit a brick wall almost immediately across multiple domains. Across all the runs, only 65% of the plans actually showed readable prices to the AI.

SPEAKER_01 2:46

Wow.

SPEAKER_00 2:47

Over a third of the time, the machine completely failed to figure out what the product cost. And on top of that, 14% of the time, the brands hid their pricing entirely behind a contact sales button.

SPEAKER_01 2:58

Okay, let's untack this. Because that contact sales button has been like the holy grail of B2B lead generation for decades.

SPEAKER_00 3:06

Oh, yeah. You gate the pricing to force a conversation.

SPEAKER_01 3:09

Right. But for an AI agent that is trying to compile a pricing matrix in 30 seconds, a contact sales button is a dead end.

SPEAKER_00 3:15

It's a complete roadblock.

SPEAKER_01 3:16

Because the agent is not going to pick up a phone. It can't fill out a form, navigate a CAPTCHA, and wait for a human sales rep to email it back on Tuesday.

SPEAKER_00 3:26

No, the AI simply abandons the task on your site. It registers your domain as unhelpful. But um the even bigger issue uncovered in that study is the JavaScript blind spot.

SPEAKER_01 3:37

Yes. The JavaScript issue. Trevor Burrus, Jr.

SPEAKER_00 3:38

Right. Because for years, the SEO industry has known that traditional Google search, you know, Googlebot has gotten incredibly proficient at rendering JavaScript. Trevor Burrus, Jr.

SPEAKER_01 3:47

It can essentially see everything a human sees on Chrome.

SPEAKER_00 3:51

Exactly. But the AI agents coming from companies like OpenAI and Anthropic, they generally do not run JavaScript when they're scanning the web for real-time answers.

SPEAKER_01 3:59

Aaron Powell, which means building a JavaScript-heavy B2B pricing page right now is basically like renting a massive, incredibly expensive billboard on the busiest highway in the world, but but choosing to paint your message in invisible link.

SPEAKER_00 4:12

That's a great way to put it.

SPEAKER_01 4:13

Like the human passengers driving by can put on special glasses to read it, but the self-driving cars, the robotic personal assistants doing the actual navigation for the buyer, they are completely blind to it.

SPEAKER_00 4:26

And to understand why this happens, you have to look at the computing mechanisms behind AI search. Loading and rendering JavaScript requires something called a headless browser.

SPEAKER_01 4:36

Which is expensive.

SPEAKER_00 4:37

Very. The server has to simulate a full visual web window, download heavy scripts, execute them, wait for third-party API calls to return the pricing data. I mean, that process takes several seconds.

SPEAKER_01 4:49

And costs real money.

SPEAKER_00 4:50

Right. It costs a fraction of a cent per page. But you multiply that fraction of a cent by hundreds of millions of user queries a day, and you literally bankrupt the AI company.

SPEAKER_01 4:59

So they just don't do it.

SPEAKER_00 5:00

Exactly. To keep costs down and speed up response times, their scrapers strip everything away and look only at the raw underlying HTML code.

SPEAKER_01 5:08

Aaron Powell, which explains the spectacular failures in the study. Like uh Zendesk, for example. Yeah. Massive platform. They had a gorgeous pricing table on their site, but it was rendered using dynamic JavaScript. When the AI agent hit that page and looked at the raw code, it couldn't find a single dollar amount. Trevor Burrus, Jr.

SPEAKER_00 5:25

It just saw blank space.

SPEAKER_01 5:26

Aaron Powell Right. So it didn't just give up on finding the price, it went to third-party blogs and review sites to guess the price.

SPEAKER_00 5:33

Aaron Powell It outsourced the narrative.

SPEAKER_01 5:35

Exactly.

SPEAKER_00 5:35

And the financial toll of inaccessible design is staggering when you look at the Databricks example from that same study. Databricks hid their pay as you go rates behind this interactive calculator.

SPEAKER_01 5:47

Aaron Powell Which an AI can't use. Right.

SPEAKER_00 5:49

The AI could not manipulate the sliders. It frustrated the agents so much that the machine entered a logic loop of failed queries. It just repeatedly tried different tactics to parse the calculator.

SPEAKER_01 5:59

Oh wow.

SPEAKER_00 6:00

Yeah. And that single failed search ultimately cost a ridiculously high 95 cents in raw API compute cost just to process.

SPEAKER_01 6:08

Aaron Ross Powell Nearly a dollar for a single failed query. I mean, the AI companies will absolutely penalize sites that burn their resources like that.

SPEAKER_00 6:15

Aaron Powell The penalty is instant abandonment. And this brings us to a fundamental concept, the context window, which is measured in tokens. Think of the AI's memory like a very small physical desk. It can only fit about 15 to 20,000 tokens, roughly 11,000 words, on that desk at any one time when it scans your page.

SPEAKER_01 6:36

That's not a lot of space.

SPEAKER_00 6:38

It isn't. So if your core facts, your pricing, your features are not sitting on that desk in plain machine readable HTML text within that strict limit, the AI cannot see them. Yep. It moves on to a competitor who did put their rates in plain text, or it relies on a review forum that might have outdated or incorrect information.

SPEAKER_01 6:57

Aaron Powell So if the AI hits that blank wall, abandons your desk, and has to find that data somewhere else, how does it choose where to go next?

SPEAKER_00 7:07

That is the big question.

SPEAKER_01 7:08

Right. It isn't just throwing a dart at the entire internet. And this leads us into the hidden pipelines of AI search. Because the network traffic analysis of ChatGPT reveals something absolutely fascinating.

SPEAKER_00 7:18

The leaked data.

SPEAKER_01 7:19

Exactly. A researcher intercepted the raw data that ChatGPT sends back to its own servers before it actually writes the response on the user screen. And hidden inside that code is a specific internal label called the result source field.

SPEAKER_00 7:34

And that field proves that ChatGPT sorts every single web source into four distinct, heavily gated pipelines. They they do not treat the internet as a flat playing field at all. The first pipeline is Tier 1, internally labeled Labrador. Think of this as the VIP room. This pipeline is exclusively reserved for major publishers, massive reference sources, and privileged content environments that often have direct, multimillion dollar licensing deals with OpenAI.

SPEAKER_01 8:01

Aaron Powell We're talking about places like Reuters, the Wall Street Journal, massive encyclopedic databases.

SPEAKER_00 8:07

Exactly.

SPEAKER_01 8:07

And to be brutally honest with the agency owners listening, your clients are not getting past the bouncer into the VIP room. You cannot buy or SEO your way into the Labrador pipeline.

SPEAKER_00 8:16

Aaron Powell No, you are locked out entirely. Which leaves tiers two and three, which are labeled Bright and Oxylabs.

SPEAKER_01 8:22

And those are actual scraping companies, right?

SPEAKER_00 8:24

Yes, those are the names of massive commercial web scraping services. This is the scraped web pipeline. When the AI needs commerce data, local business info, or structured facts about a B2B service, it sends these commercial scrapers out to grab the raw HTML.

SPEAKER_01 8:39

Okay.

SPEAKER_00 8:40

This is the arena where 99% of business websites actually compete.

SPEAKER_01 8:45

Wait, so if I'm a CMO or an agency owner listening to this, does this mean my expensive thought leadership content strategy is useless if I'm not in that VIP room? Yeah. Like, do we just fire our front-end designers and strip our beautiful websites down to plain HTML text like it's 1995?

SPEAKER_00 9:01

Aaron Powell Well, if we connect this to the bigger picture, it doesn't mean abandoning good design. No. But it absolutely demands a shift in priorities. We are moving from traditional keyword management to data architecture. Trevor Burrus, Jr.

SPEAKER_01 9:12

Okay, data architecture.

SPEAKER_00 9:13

Right. In that scraped tier of the Brighton Oxylabs pipelines, the brutal truth is that your visibility depends entirely on whether a machine scraper can extract your core facts faster and more clearly than your competitors.

SPEAKER_01 9:24

So the foundation has to be solid.

SPEAKER_00 9:26

Exactly. Your site can have a beautiful interactive paint job for the human users, but the engine underneath the code that loads first must be pure, universally accessible HTML text.

SPEAKER_01 9:38

Got it. But uh if the first three tiers are VIPs and commercial scrapers, there has to be a fallback, right? Like what happens if the AI just needs to do a general web query? Is traditional Google or Bing completely ignored now?

SPEAKER_00 9:50

Aaron Powell No, the fallback is tier four, labeled SERP. This is essentially traditional search engine infrastructure or like Microsoft Bing's index layered directly into the AI as a baseline search.

SPEAKER_01 10:01

Aaron Powell Okay, so SEO isn't dead?

SPEAKER_00 10:03

Aaron Powell No, traditional SEO signals and backlinks still have some role, but they're just one piece of a much larger, much more complex routing puzzle.

SPEAKER_01 10:11

Aaron Powell So let's say an agency has done everything right up to this point. They've optimized for that scraped web tier, their pricing is in plain text, their data architecture is flawless, the headless browser issue is solved, and the AI can read the site perfectly. Are they the one who gets recommended to the buyer?

SPEAKER_00 10:26

See, that is the trap. The assumption that being readable guarantees being recommended is a massive trap for marketing teams. We need to completely redefine what visibility actually means in an AI-driven world.

SPEAKER_01 10:40

Because being scraped and being trusted are two entirely different metrics.

SPEAKER_00 10:44

Completely different.

SPEAKER_01 10:45

Because according to the data, there are three distinct states a brand can occupy inside an AI's brain like ChatGPT. You can be fetched, you can be cited, or you can be mentioned.

SPEAKER_00 10:57

Right. Let's break those down. Fetched simply means your page was pulled into the AI model's working memory. It looked at your site.

SPEAKER_01 11:04

That's step one.

SPEAKER_00 11:06

Exactly. Cited means your page was actually used as the clickable footnote, the authoritative proof behind a specific claim the AI made to the user.

SPEAKER_01 11:14

And mentioned.

SPEAKER_00 11:15

Mentioned means your brand name appeared in the final output, but someone else's website was cited as the proof for why your brand is good.

SPEAKER_01 11:22

Here's where it gets really interesting. Let's look at the data comparing how ChatGPT treats Reddit versus YouTube.

SPEAKER_00 11:28

This was eye-opening. It's crazy. The network analysis showed that YouTube pages were fetched 201 times by the AI during a testing sequence. Like the AI pulled the data into its memory 201 times. Right. But out of all those fetches, the number of times a YouTube video was actually cited as the source of truth was zero. Absolutely zero.

SPEAKER_01 11:51

Contrast that with Reddit. Reddit threads were fetched 278 times, and they were cited 11 times.

SPEAKER_00 11:58

So why there's a huge difference?

SPEAKER_01 12:00

To understand the disparity, you have to look at the mechanics of how the machine reads. A YouTube page is overwhelmingly video content, right? It only hands over surface level metadata to a scraper.

SPEAKER_00 12:11

Titles, tags, maybe a brief description.

SPEAKER_01 12:14

Exactly. An AI agent cannot watch the video to verify a complex claim like this software integrates seamlessly with Salesforce. There is no text transcript for the AI to grout onto. But a Reddit thread, however, is pure unadulterated text. A citation has to latch onto a real verifiable sentence. And this highlights the exact trap agencies are falling into. If you are an agency owner and you're generating monthly reports for your clients that boast about how many times their brand was mentioned by ChatGPT or how many times their pages were fetched, you are selling fools gold.

SPEAKER_00 12:48

Yeah, those are vanity metrics now.

SPEAKER_01 12:50

Being present in the AI's working memory is not the same as being trusted. If your client's website isn't the citation, they're not the authority in the room.

SPEAKER_00 12:59

Earning that citation requires understanding how aggressive the AI is when it evaluates competing products. When Chat GPT is doing a deep comparison for a buyer, it doesn't just run one simple search.

SPEAKER_01 13:11

It goes deep.

SPEAKER_00 13:12

Very. It might fan out into 40 different subqueries in the background. It will literally execute a command like cite callinyourdomain.com slash pricing just to probe your site's specific architecture.

SPEAKER_01 13:24

Just hunting for specific detail.

SPEAKER_00 13:25

Exactly. It scans the raw HTML specifically hunting for a dollar sign or a euro symbol to verify the price it thinks you charge.

SPEAKER_01 13:32

So it's not just casually reading your marketing brochure, it's acting like a forensic accountant auditing your receipts. It is scanning purely for the verification symbol.

SPEAKER_00 13:40

And the machine is also ruthless about deduplication.

SPEAKER_01 13:44

Oh, right.

SPEAKER_00 13:44

In traditional SEO, agencies used to spin up 20 thin, highly repetitive blog posts targeting slight variations of a keyword to dominate search results.

SPEAKER_01 13:54

The classic spam strategy.

SPEAKER_00 13:55

Right. But if you try that with an AI agent, it identifies the repetitive overlap and collapses all 20 pages into a single domain entity in its memory. You cannot spam an AI with sheer volume.

SPEAKER_01 14:09

So what does this all mean for a content strategy moving forward? Like if you can't overwhelm the machine with volume on your own blog, how do you win the citation?

SPEAKER_00 14:17

Well, you cannot just cite yourself. An AI evaluates your own marketing copy with extreme skepticism.

SPEAKER_01 14:23

Of course.

SPEAKER_00 14:24

It synthesizes its final verdict about your brand using other people's words. If it concludes that your software is budget-friendly or the most popular choice, it is pulling those exact labels from third-party reviews, industry forums, and independent comparison sites.

SPEAKER_01 14:38

So the strategy has to shift off-site.

SPEAKER_00 14:41

Yes. Your agency needs to pivot heavily toward earning third-party proof and reputation management across the wider web. Publishing endless self-congratulatory blog posts on your own domain. I mean, it just isn't enough anymore. You need the rest of the internet to validate your claims in plain HTML text.

SPEAKER_01 14:58

Wow. Okay, so we've been building this entire strategic discussion around the idea that the AI is actually searching the live web to find these answers. Right. But our sources reveal a scenario where the AI doesn't even bother looking at the live web at all.

SPEAKER_00 15:11

This is a huge catch. Hidden in that same network traffic data is a routing parameter called the turn use case bucket.

SPEAKER_01 15:17

The turn use case.

SPEAKER_00 15:18

Right. Before ChatGPT attempts to launch a commercial scraper or check the traditional search index, it categorizes the user's prompt. Sometimes, based purely on the phrasing, it drops the prompt into a bucket labeled simply text.

SPEAKER_01 15:32

And when it categorizes a query as text, it performs zero live web searches. None. No scrapers, no VIP pipeline. It relies entirely on its internal training data, the stale historical memory baked into the model months or years ago.

SPEAKER_00 15:48

And the researcher who discovered this gave a chilling example for anyone working in, say, healthcare or pharmaceutical marketing.

SPEAKER_01 15:56

Yeah, this was scary.

SPEAKER_00 15:57

They asked ChatGPT for the latest treatment guidelines for type 2 diabetes. Now, you would reasonably expect an AI to scour the web for the absolute most recent peer-reviews medical journals.

SPEAKER_01 16:09

Obviously, it's medical advice.

SPEAKER_00 16:11

Instead, the system categorized the prompt as text and answered entirely from its static memory without performing a single live web search.

SPEAKER_01 16:19

So the exact wording of the user's prompt is what pulls the levers behind the curtain. Like if a user types best 4K TVs to buy, the AI interprets the commercial intent and triggers a live shopping search using scrapers. Yes. But if they type best 4K TVs with reviews, it triggers a standard non-commercial web search. It's the exact same product category, but a completely different machine pipeline.

SPEAKER_00 16:41

And while B2B brands are out there struggling with JavaScript blind spots and begging scrapers to read their pricing pages, e-commerce brands have been quietly handed a golden ticket. There is a VIP shortcut built specifically for retail.

SPEAKER_01 16:55

But wait, if plain HTML text is the absolute rule for readability, how are massive e-commerce sites with thousands of dynamic, heavily scripted product pages surviving?

SPEAKER_00 17:06

Yeah.

SPEAKER_01 17:06

Because those sites are entirely driven by complex JavaScript.

SPEAKER_00 17:10

They survive because OpenAI has built direct data pathways for commerce. Things like direct Shopify catalog integrations and raw product feeds. If you're an e-commerce brand, you don't have to cross your fingers and pray that the bright scraper correctly renders the complex script on your product page.

SPEAKER_01 17:27

You just bypass it.

SPEAKER_00 17:28

You bypass the visual web entirely. You can pipe your structured product data, your real-time inventory, and your exact pricing directly into the AI's database via an API handshake.

SPEAKER_01 17:37

And with ChatGPT and other agents moving rapidly toward instant in-chat checkouts, the distance between discovering a product and buying it is completely vanishing.

SPEAKER_00 17:46

It's immediate.

SPEAKER_01 17:47

Discoverability and the actual transaction are collapsing into a single chat-based interface. I mean, I have to ask you, the listener, a very direct question right now. If ChatGPT had to explain your business category to a buyer today, using only its historical memory or a plain text scrape of your website, would you be a primary trusted source? Or would you be a complete afterthought, losing the sale to a competitor who simply formatted their data better?

SPEAKER_00 18:14

This raises an important question about resource allocation for anyone listening. The action plan here is incredibly clear for agencies and business owners.

SPEAKER_01 18:23

What's step one?

SPEAKER_00 18:24

First, audit your machine readability immediately. Have your development team turn off JavaScript in a testing browser and see if your pricing, your core features, and your value proposition instantly disappear.

SPEAKER_01 18:34

If they do, you're in trouble.

SPEAKER_00 18:35

Big trouble. Second, aggressively map out your third-party citation environment. Where is the AI going to find independent text proving that you are good at what you do?

SPEAKER_01 18:45

You need that outside validation.

SPEAKER_00 18:46

Absolutely. And third, if you are in the e-commerce space, you must start treating structured product feeds as frontline media infrastructure. It is no longer just back office plumbing relegated to the IT team. It is your primary, highest converting marketing channel.

SPEAKER_01 19:02

So what does this all mean as we look toward the future of search? The main takeaway is that Chat GPT and its competitors are not pulling from one single unified equal internet.

SPEAKER_00 19:13

No, they're not.

SPEAKER_01 19:14

They're pulling from multiple deeply hidden source economies, and each one has its own rigid rules, its own computing costs, and gatekeepers. The winners in this new era will not be the brands that publish the highest volume of content or the ones with the flashiest, most interactive websites.

SPEAKER_00 19:29

Not anymore.

SPEAKER_01 19:30

The winners will be the organizations that make their facts the absolute easiest for machines to retrieve, verify, and trust.

SPEAKER_00 19:37

It really represents a fundamental, irreversible shift from human persuasion to machine integration.

SPEAKER_01 19:43

And I want to leave you with a final provocative thought to mull over today. We've talked extensively about how AI agents are already evaluating your features and pricing on behalf of human buyers.

SPEAKER_00 19:53

Right.

SPEAKER_01 19:53

We've talked about the rise of instant in-chat checkouts where the AI facilitates the sale. But take that trajectory one step further. If the buyer is using an AI to find the best price, and you are using structured data to feed your pricing directly into the machine, how long until we see a buyer's AI agent actively negotiating pricing and contract terms directly with your company's AI?

SPEAKER_00 20:18

That changes everything.

SPEAKER_01 20:19

Imagine a complex B2B transaction where no human marketing or sales team is involved on either side. Are you prepared for a business world that is entirely machine to machine?

SPEAKER_00 20:30

And honestly, the technology to facilitate that negotiation is already being built. It is coming significantly faster than the market realizes.

SPEAKER_01 20:37

Well, thank you for joining us on this deep dive. We'll see you next time as we continue to unpack the rapidly changing machinery of digital business.