Retrieval Optimization Applied to Two Brands (One Big, One Small)

Posts 10 (“Retrieval Optimization: It’s Not Just SEO or GEO“) and 11 (“Retrieval Optimization at Each Layer“) built the framework. Four layers — L1 Entity Establishment, L2 Entity Depth, L3 Category Citation, L4 Content Optimization — with SEO and AI visibility as two lenses reading the same system.

This post stress-tests it. Same framework applied to two real brands at opposite ends of the spectrum: one global crypto exchange with a decade of structural press coverage, one regional web design agency going national with 19 years of operation and almost no earned media footprint. Both are Loganix clients.

The point isn’t just “here’s how the layers look for each brand.” The point is that the same four layers produce completely different maps, completely different gaps, and completely different build orders depending on where the brand sits. That’s what makes the framework useful rather than theoretical — it doesn’t give you the same answer twice.

Everything below comes from publicly observable information. You could run the same exercise on any brand with a browser.

Brand 1 — Kraken

The crypto exchange whose name happens to match the Seattle hockey team. I’ll drop the wink because most of you have figured it out by now and because the name has some relevance in how things play out for them.

Founded in 2011, 10M+ clients globally, $665B in 2024 trading volume, first crypto exchange to receive a U.S. bank charter, decade of major financial press, Wyoming SPDI approval, MiCA regulatory approval in mid-2025, and SEC litigation dropped in early 2025. Basically, the established-brand version of this exercise.

L1 — Entity Establishment

Publicly observable: Strong with a small AI-side opportunity.

Kraken is a well-established entity by any measure. Wikipedia entry, Wikidata entry, Crunchbase, LinkedIn at 278K followers, every major directory, schema on the main pages. From the SEO lens, L1 looks solved. Google has them in its Knowledge Graph; NAP is consistent, structured data is in place, and the entity has been unambiguous in Google’s view for years.

From the AI lens, L1 is also solved, but the disambiguation layer is more interesting. ChatGPT, Gemini, Claude, and Perplexity all know who Kraken is without retrieval. But the brand name collides with the Seattle Kraken hockey team, various ships, mythological references, and a handful of smaller companies with the same name. Kraken has the scale and the training data weight to override this, but there’s still a small amount of disambiguation work happening at query time that ideally would be done at the source level instead.

The retrieval optimization move at L1 here isn’t structural. The schema is fine. It’s making sure that every public-facing description pairs the brand name with the category (“cryptocurrency exchange”) in close proximity, so the AI lens is reading brand-to-category co-occurrence as a primary training signal instead of resolving ambiguity at query time. Same source, slightly different read, small optimization that costs nothing.

L2 — Entity Depth

Publicly observable: Structurally strong from historical press. The maintenance side requires data we can’t see.

L2 is where Kraken’s decade of press coverage becomes obvious as a retrieval optimization asset, even though nobody called it that at the time.

Bloomberg, Reuters, Wall Street Journal, CNBC, TechCrunch, Forbes — continuously for over a decade. The Mt. Gox bankruptcy story alone gave them a multi-year reputational anchor with major financial press. Being the first crypto exchange with a U.S. bank charter gave them another. MiCA approval gave them another. Wyoming SPDI gave them another. None of this is small coverage. It’s structural press presence in the highest-authority financial outlets.

From the SEO lens, that press history has been doing link equity work for a decade. Referring domains in the tens of thousands, anchor text diversity, and Domain Rating are where they should be. Standard strong L2 SEO footprint.

From the AI lens, the picture is stronger in a different way. All that press coverage is what put Kraken into the training data of every major LLM at depth. The brand-to-attribute associations are rich because the press coverage has been rich for years. ChatGPT knows Kraken is “one of the longest-running U.S. crypto exchanges.” Claude knows about the Wyoming bank charter. Gemini knows about Mt. Gox. The model knows what Kraken is associated with because the coverage has consistently told it what Kraken is associated with. That’s L2 retrieval optimization paying off without anyone labeling it that way.

The part I can’t assess from a public read: the maintenance side. Training-layer signals come from historical press. Retrieval-layer signals come from fresh press. The AI lens pulls the freshest sources first when generating answers about the current state. You can’t tell from outside whether the cadence and brief format are serving the AI lens correctly right now, and that’s the kind of question the deeper diagnostic work would surface.

The framing point this illustrates: at L2, the SEO lens and the AI lens can both read strongly while pointing at different things. SEO reads link equity. AI reads brand-to-attribute associations in training data. Same press placements, different cargo extracted.

L3 — Category Citation

Publicly observable: Genuinely strong and verifiable in real time.

L3 is the layer where you can do the most useful public verification, and it’s also the layer where the work Kraken has done becomes most visible.

The category question at L3 is simple: when someone asks an AI, “What’s the best crypto exchange,” does Kraken show up?

The answer is yes, repeatedly, across most of the major platforms. And the reason you can observe this from outside is that the editorial coverage of “best crypto exchanges” listicles is dense with Kraken mentions. I just ran a search for “best crypto exchanges 2026” and saw Money.com naming them “Best Crypto Exchange Overall for March 2026.” Koinly lists them in their top 10 for U.S. exchanges. Yahoo Finance features them. Most of the major comparison content puts Kraken in the top 3-5 of any “best of” list in the category.

From the SEO lens, this is standard digital PR — editorial mentions on high-authority sites producing link equity and brand authority for category queries.

From the AI lens, it’s more consequential. ChatGPT and Gemini answer “best crypto exchange” by scanning exactly these editorial listicles. The Money.com piece, the Koinly piece, the Yahoo Finance comparison — these are the sources retrieved when someone asks an LLM the category question. If a brand isn’t in those listicles, it’s invisible to the AI for that query, regardless of organic ranking for the same terms.

Kraken is in those listicles consistently. That’s L3 retrieval optimization doing its job in real time. You can verify it yourself in five minutes.

Two things worth pulling out:

The brand has to be a credible top-3 candidate before any of this works. You can’t push your way into “best crypto exchange” lists from nothing. The L3 work compounds on a foundation that has to exist first. Kraken has the foundation (security history, longevity, scale, regulatory wins) AND the L3 placement strategy. Both matter.

L3 has multiple sub-pools per query. “Best crypto exchange” is one query. “Best crypto exchange for day trading” is another. “Lowest fee crypto exchange” is another. Each has its own listicle landscape and its own retrieval pool. A brand can be strong in one sub-pool and absent from another, and the actual per-sub-pool audit is the kind of thing the deeper diagnostic work handles.

L4 — Content Optimization

Publicly observable: Decent SEO maturity, typical commercial-page opportunity.

L4 is the on-site content layer. Are the pages on kraken.com structured for retrieval by both Google and AI? Unlike the earlier layers where I was mostly working from signals OUTSIDE kraken.com, this layer I can read directly by loading the pages and looking.

A few specific observations from the public read:

The fee schedule page (kraken.com/features/fee-schedule) has a solid tabbed structure separating Spot Crypto, Stablecoins, Futures, Margin, and other fee types. The actual numbers exist on the page in tables. But the opening copy on each tab is soft positioning (“With our competitive stablecoin fee structure, we aim to deliver deep liquidity, tight spreads, and minimal slippage”) rather than answer-first fact delivery. An AI trying to extract “what does Kraken charge for spot crypto” has to work through the positioning copy to get to the tables. An answer-first version would open with the fee range directly (“Spot crypto maker fees start at 0.00% and taper to 0.25% based on 30-day trading volume”) before any positioning language. The facts are all there — they’re just not front-loaded.

The supported currencies page uses a list format that’s scannable for humans but doesn’t make extractable claims. And here’s the interesting thing — if you pull the number of supported assets from different sources, you get wildly different answers. CoinMarketCap says 120 coins. CoinCodex says 180+. Kraken’s own LinkedIn says 200+. Money.com‘s 2026 review says 600+. Contrary Research’s company profile says 531 digital assets as of October 2025. Kraken’s own learn content says “over 500 digital assets” in one place and “over 480 cryptocurrencies” in another. Six different sources, six different numbers, all citing Kraken.

That variance isn’t an AI limitation — it’s a first-party content quality issue. The authoritative answer to “how many assets does Kraken support” should be one number, owned by kraken.com, updated continuously, and marked up so retrieval systems can point at it cleanly. Right now, the answer depends on which source the AI happens to retrieve, and different platforms are likely citing different numbers for the same question. That’s an L4 retrieval optimization opportunity that surfaces directly.

Fee numbers have similar variance across the public record. Third-party coverage cites maker fees of 0.00% to 0.16% (CoinMarketCap), 0.00% to 0.25% (Kraken’s own learn content), and 0.25% to 0.40% (Koinly), depending on the source and the product being referenced. Some of this is legitimate product variation (Pro vs Instant Buy vs Stablecoin vs Futures). Some of it is stale third-party data that Kraken has no way to correct. The L4 move here is making the authoritative structured answer on kraken.com so clean and machine-readable that the third-party variance matters less, because the retrieval system can find the canonical source directly.

The product pages themselves — I loaded the main features page and the product category pages. They’re well-designed, clean, and brand-consistent. But they share a pattern with most financial services sites: the opening content is brand positioning, the commercial facts are buried below the fold or in secondary pages, and the structural emphasis is on visual hierarchy for humans rather than extraction hierarchy for AI. This isn’t a broken L4; it’s an L4 that was optimized for one reader when it could be optimized for both.

What I genuinely can’t see from a public read:

• Whether any of the L4 work is actually moving citation rates across AI platforms (which requires citation tracking across 5+ platforms over time)
• Whether the same patterns hold across the thousands of other pages on the site that I didn’t spot-check
• Whether the commercial pages are getting the freshness cadence their category requires
• Which specific structural choices are correlating with which specific AI platform citation behavior

Those are the longitudinal and at-scale questions the deeper diagnostic work handles. The spot-check observations above are what anyone can do in an hour with a browser.

The general L4 pattern at Kraken: the facts exist, the structure is competent, and the opportunity is front-loading the commercial answers on the highest-value pages (fee schedule, supported assets, product features) so that AI retrieval finds the canonical first-party answer instead of pulling variance from the third-party ecosystem. GPT-5.4 already sends 56% of its citations to first-party brand websites per the Writesonic study — the model wants to cite kraken.com directly for these claims. The L4 RO work is making those pages easier to cite cleanly.

Brand 2 — Bizango

Now for the other end of the spectrum. Bizango is a Seattle-based web design and branding agency founded in 2006 by Mark Figlozzi and Susan Jackson. They are a seven-person team across Seattle, Colorado, and New Mexico. 19 years of operation. Real client portfolio including AWS, Keller Rohrback, Jackson Remodeling, and multiple architecture and construction firms. Starting project prices around $8-30K. Real Yelp listing, real Clutch profile, real reviews, real industry specialization pages for attorneys, construction, architects, engineers, tourism, and ecommerce.

Same time horizon as Kraken (both founded 2006-2011). Completely different retrieval optimization position.

L1 — Entity Establishment

Publicly observable: Decent foundation. Real disambiguation issue. Geographic framing mismatch.

The business entity is established. They exist, they have a real address, phone number, founding date, founder bio, team of 7, Clutch profile, Yelp listing, client portfolio, and reviews. From the SEO lens, L1 is fine as a baseline.

But there are two observations worth flagging.

First, the name. “Bizango” isn’t actually a unique brand word — it’s a Haitian Vodou term referring to secret societies, with cultural and religious meaning that shows up in training data alongside the agency. Kraken has a similar collision (hockey team, mythology), but Kraken has the scale and training data weight to override it. Bizango doesn’t. When an AI platform encounters “Bizango” without strong contextual anchoring, it’s surfacing a word that has two distinct meanings in training data, and the cultural reference likely has more presence than the agency. That’s a meaningful L1 disambiguation problem that shows up almost entirely in the AI lens — Google has long since resolved it through Knowledge Graph and link signals.

Second, the geographic framing. The brand is consistently called “Seattle web design company” across every page of the site, and Seattle appears in the H1 of the home page. But the team is actually in three states (Seattle, Colorado, and New Mexico), and they serve clients nationally. From the SEO lens, the Seattle framing pays off in local search — they rank for “Seattle web design” terms. From the AI lens, they’re harder to surface for non-Seattle queries because the entity has been explicitly geographically anchored to one city despite serving nationally. A prospect in Denver asking an AI for a web design agency for their construction company is less likely to get Bizango surfaced than they should be, given that the team actually lives in Denver’s time zone.

The L1 retrieval optimization observation here: both of these are invisible to the SEO lens (Google has solved disambiguation through other signals, and the Seattle framing is a local SEO strength), but they’re meaningful drags on the AI lens. You can only see them if you’re looking through both lenses at the same time.

L2 — Entity Depth

Publicly observable: Thin. This is the structural gap.

This is where Bizango’s picture diverges hardest from Kraken’s, and it’s the most important thing this side-by-side comparison surfaces.

Bizango has been operating for 19 years. That’s the same time horizon as Kraken. In theory they’ve had the same amount of time to build the kind of third-party authoritative coverage that creates training-layer brand presence in LLMs. In practice, the picture is completely different.

What I can find publicly:

• Yelp listing (directory)
• Clutch profile (directory)
• UpCity profile (directory)
• ReportGarden profile (directory)
• Goodfirms listing (directory)
• DesignRush listing (directory)
• Sortlist listing (directory)
• Their own website, blog, and reviews page
• Client testimonials on their site

What I can’t find:

• Coverage in major design publications (Smashing Magazine, A List Apart, Awwwards, Webby Awards)
• Features in Seattle business press (GeekWire, Puget Sound Business Journal, Seattle Times business section, Seattle Met)
• Speaker slots at design or marketing conferences with public coverage
• Industry awards beyond directory listings
• Authoritative third-party mentions that aren’t paid directory placements
• Authored articles in design trade publications

That’s the L2 gap, and it’s structural. They have 19 years of real client work, but it hasn’t translated into the kind of earned media coverage that builds training-layer brand-to-attribute associations in LLMs. From the SEO lens, they probably have backlinks from directories and clients, but limited high-authority editorial backlinks. From the AI lens, the model doesn’t really know who Bizango is at depth. There’s no “Bizango is known for X” pattern in training data because the press to create that pattern doesn’t exist.

This is the most important single observation in this entire analysis: two brands with the same time horizon, wildly different L2 positions. Kraken got the press by being in a category that generated press for them. Bizango operates in a category where press has to be actively earned by PR effort, and they haven’t done that work, or haven’t done enough of it.

The SEO lens reads this as “weaker backlink profile than the category leaders.” The AI lens reads this as “the model barely knows this brand exists at depth.” Same gap, two different consequences, both of which compound.

L3 — Category Citation

Publicly observable: Split. Strong in one half, weak in the other.

L3 is where Bizango’s picture gets interesting because it’s not uniformly weak — it’s split down the middle in a way that teaches something specific about how L3 actually works.

The good news on L3: Bizango shows up in the major directory-driven listicles for Seattle web design. Clutch’s “Top Web Design Companies in Seattle – March 2026” ranks them at the top of the list. Goodfirms lists them. UpCity lists them. DesignRush lists them. Sortlist lists them. That’s real L3 presence on directory-style listicles that AI platforms scan heavily for category queries.

The complicated news: the non-directory editorial listicles tell a completely different story.

• Mojo Media’s “Best Web Design Agencies in Seattle (2026)” (the editorial think-piece version) doesn’t mention Bizango. They feature Culture Foundry, Seattle New Media, and others.
• Fueler.io‘s “Top 10 Web Design Agencies in Seattle (2026)” features efelle creative, Thrive Design, Sayenko Design, and UPQODE. Bizango isn’t in the top 10.
• The actual editorial content, where humans wrote opinions about Seattle agencies, tends to list different names.

This split matters because directories and editorial content get treated differently by AI platforms. ChatGPT scans both but weights them differently depending on the query type. For “best web design agency Seattle,” ChatGPT might pull from Clutch first. For “I’m a law firm, and I need a web designer who understands my category, what agencies should I consider?” ChatGPT is more likely to pull from editorial content that has opinions and reasoning behind recommendations. And that’s where Bizango’s L3 position is weaker.

There’s one more L3 layer worth calling out specifically. Bizango has dedicated industry pages — attorneys, construction, architects, engineers, tourism, and ecommerce. That’s an explicit category play. The question the framing surfaces: when someone asks ChatGPT “best web design agency for law firms” or “best web design agency for construction companies,” are those industry pages actually surfacing Bizango in the retrieval pool for those sub-category queries? A quick public check suggests they rank for some of these terms in Google, but don’t appear in the top editorial content written about web design for those specific industries. The pages exist and do SEO work. The per-sub-pool L3 retrieval position of those pages is the open question.

The L3 observation of the framing makes visible: Bizango has real category presence in the directory half of L3 and a meaningful gap in the editorial half of L3. That’s not a single problem with one fix — it’s two different problems that require two different solutions. The directory presence is already working. The editorial listicle presence needs to be earned through a fundamentally different kind of work (original content, opinion pieces, public case studies, design community engagement, speaker placements, commentary on design trends) than what gets you into Clutch’s ranking.

L4 — Content Optimization

Publicly observable: Classic agency website pattern.

From a public read of their main pages, the L4 picture is the one 90% of agency websites have. It’s recognizable because it’s so common.

The good:

• Clean modern design, mobile responsive
• Schema markup present
• Internal linking between main pages and industry pages
• Real portfolio with real case studies
• Dedicated reviews page with real client testimonials
• FAQ-style content on the web design services page
• Content hierarchy is scannable

The gaps from an AI extraction perspective:

• Heavy emotional/visual messaging in opening paragraphs. “Inspire confidence. Win new customers. Play to your strengths. Does your website work for you?” — beautiful copy, but not answer-first. An AI trying to extract “what does Bizango do and what makes them different” isn’t getting clean structured information from that opening.
• Service descriptions are aspirational rather than specific. “We empower creative marketers and help them tell their stories with flexible tools that put them in the driver’s seat” is evocative but doesn’t give an AI anything extractable about services, deliverables, timeline, or approach.
• Pricing information exists, but it’s buried in an FAQ on a sub-page, not directly extractable from the main service pages. When an AI is asked, “How much does Bizango charge?” the answer is harder to retrieve than it needs to be.
• Industry pages exist, but the content on them is closer to marketing copy for the main service than industry-specific answer-first content. A law firm considering Bizango wants to see specific examples, specific approaches, specific pricing, and specific outcomes for law firms. The current pages give them emotional positioning for Bizango generally.
• The client logo wall is great for human trust signals, but doesn’t give an AI extractable structured information about who Bizango actually works with.

This is exactly the L4 pattern Post 11 described as “the signals that get content cited by AI are the same signals that make content perform well in Google, with small platform-specific tuning at the edges.” Bizango’s L4 issues are the same issues you’d flag in a standard SEO content audit focused on answer-first structure. The framing doesn’t invent new problems here — it just makes the existing problems visible through a second lens as well as the first.

The L4 framing point: the gaps aren’t broken design or broken copywriting. The site is well-crafted for a human reader to have an emotional response. It just isn’t optimized to be read by an AI trying to extract specific factual claims about what the agency does, who it serves, and what it charges. Both are legitimate goals, but one of them isn’t being served right now.

Cross-comparison: the same framework, two different maps

Looking at the two brands side by side, a few patterns emerge that are more instructive than either brand in isolation.

Same time horizon, radically different L2 positions. Both brands have been operating since 2006-2011. Kraken has built a structural press presence with the highest-authority financial outlets continuously. Bizango has run directory listings and collected client testimonials. Both are valid choices a founder could make, and both took 19 years to produce. But from a retrieval optimization perspective, they’re completely different positions. Kraken’s L2 is doing massive training-layer work for the AI lens. Bizango’s L2 is doing almost nothing.

Both have L1 disambiguation issues. Only one has the scale to absorb them. Kraken and Bizango both have brand names that collide with other things in training data (Kraken with hockey and mythology, Bizango with a Vodou term). Kraken has the scale to override this through sheer weight of brand-category co-occurrence in training data. Bizango doesn’t, which means the disambiguation issue has real retrieval consequences for them that it doesn’t have for Kraken.

The lenses disagree about which layer is the strongest, in different ways for each brand. Kraken’s SEO lens reads L4 as mature. The AI lens reads L3 as the most visibly working layer. For Bizango, the SEO lens reads L3 as strong (they rank for category terms and show up in directories). The AI lens reads L3 as split because the editorial half is mostly missing. Same framework, both brands, four different lens readings producing four different “where are you strongest” answers.

The build order question is real for one brand and not for the other. Kraken has all four layers working at some level. The question for them is incremental optimization and cadence maintenance. Bizango has dependencies between layers that matter: L2 thinness limits L3 retrieval (no rich training-layer brand-to-category association for the AI to pull from), L1 disambiguation compounds L2 thinness (even when their name surfaces, the AI isn’t sure which Bizango), and L4 gaps on the commercial pages compound the category presence weakness at L3. The layers aren’t independent — they feed each other.

For Bizango, the framing suggests a specific build order that would matter more than picking any single layer to fix. L1 disambiguation first (explicit category and geographic clarification in every public-facing description), then L2 press work to build the training-layer brand-to-attribute associations that L3 editorial coverage needs, then L3 editorial placement through commentary, speaking, and public case studies, then L4 page-level restructuring to serve both lenses cleanly on the commercial pages that matter most. That’s the dependency order. Fixing L4 first without L2 in place doesn’t get the AI to cite pages that exist but don’t exist in the AI’s category consideration set. Fixing L3 editorial without L2 training-layer work is hard because editorial writers pull from the same training data the AI does when they decide who to feature.

None of these observations is unique to Bizango. The same pattern shows up in most mid-sized service businesses that have decent SEO maturity but haven’t done the earned media work that builds training-layer presence. What the framing does is make the sequence and the dependencies visible.

What this exercise is and isn’t

This is a public-information walkthrough, not the full diagnostic. A real retrieval optimization audit runs on data you can’t observe from outside — citation tracking across platforms, page-level structural analysis, PR cadence patterns, sub-pool category presence per query, longitudinal change measurement. That’s the work we’re building products around. What you just read is what the framing surfaces with nothing but a browser and careful searching.

Two things it makes visible that neither the SEO lens nor the AI lens alone would:

First, the layers have dependencies. Bizango’s L2 thinness limits L3 retrieval. L1 disambiguation compounds L2. L4 gaps compound L3 weakness. Build order matters, and the framework shows you which order.

Second, work that’s already paying off isn’t always labeled as retrieval optimization. Kraken’s press history is “brand building” in C-suite language and “digital PR” in agency language. The framing explains why it’s working and what to protect.

More layer-specific posts coming next, with tactical specs at each level.

This article was originally published on X by Aaron Haynes. Aaron is the CEO of Loganix, a visibility + SEO platform for brands and agencies.