Making Complex Products Understandable

The problem we exist to solve

The best product in a category does not always win. Often the winner is the product buyers understand.

That gap is widest for complex products: a new processor architecture, a network fabric, an energy-management platform, a developer API that replaces a team. The people who build them know exactly why they matter. The people who buy them have to piece that together from fragments, on a deadline, often before they ever talk to the builder.

More and more often, the first explanation a buyer hears comes from an AI answer engine. Someone asks ChatGPT, Gemini, Google AI Overviews, Perplexity, or Claude what a product does, how it compares, and whether it fits, and the answer becomes their first working model of the product. If that answer is missing, vague, outdated, or wrong, the product can lose the buyer before anyone on its team knew the buyer existed.

SolCrys exists to make complex products understandable to the people who need them. Today we start where more and more buyers start, with answer engines. Our discipline is Answer Engine Optimization, and our method is the Loop. We measure how answer engines explain a product, diagnose where the explanation breaks, execute fixes grounded in the company's own approved facts, and verify that the explanation changed. Then we run it again.

Why complex products get misunderstood

Complex products are rarely misunderstood because their makers are careless. They are misunderstood because the people who know the most are badly placed to see what a newcomer is missing, and because buyers now do most of their learning without the maker in the room.

The builder's knowledge is the first obstacle. Economists call it the curse of knowledge. Camerer, Loewenstein, and Weber (1989) found that "better-informed agents are unable to ignore private information even when it is in their interest to do so." In their introduction, they name the case that matters here: "sellers are better informed about the true value of their products than buyers."

In Elizabeth Newton's 1990 Stanford study, people tapped out well-known songs and predicted listeners would name about half of them. Listeners named 3 of 120, or 2.5%. Every engineer who has written a product page has been the tapper.

The buyer's situation makes it worse. A purchase is now a committee decision made mostly at a distance:

• A typical B2B buying decision now involves 13 internal stakeholders and 9 external influencers, according to Forrester's State of Business Buying, 2026.
• B2B buyers spend only 17% of their purchase process talking with potential suppliers, writes Gartner's Brent Adamson.
• Technical buyers complete 62% of their journey online before contacting a vendor, and 21% say the technical complexity of a solution is what made them reach out (GlobalSpec, TREW Marketing, and Elektor, 2026).

So a complex product has to be understood by a group of people, most of whom will never meet the people who built it.

Confusion rarely ends in a no. It ends in nothing. In a study of more than 2.5 million recorded sales conversations, Matthew Dixon and Ted McKenna found that 40% to 60% of deals are lost to customers who intended to buy and then failed to act. Forrester reported in 2024 that 86% of B2B purchases stall.

Lab research shows how little it takes. When the same two products were described in a hard-to-read font, the share of people who put off choosing rose from 17% to 41% (Novemsky, Dhar, Schwarz, and Simonson, 2007). In a 650-person experiment, technical jargon led to "increased risk perceptions, and lower support for technology adoption" (Bullock and colleagues, 2019).

The pattern is consistent. What a buyer cannot understand feels risky, and what feels risky gets postponed. For a complex product, understanding is not a communications nicety: it is the step between being found and being chosen, and it is where much of the value leaks out.

Why now: the explainer changed

For 25 years, a company's website was the authoritative explanation of its product, and search engines sent buyers to it. Now the explanation often happens inside the answer, before the buyer visits anything.

The habit formed in about two years. Across 6 countries, the share of people who used generative AI to get information in the past week rose from 11% in 2024 to 24% in 2025 and 38% in 2026 (Reuters Institute, October 2026). Google says AI Overviews has more than 2.5 billion monthly users, and the Gemini app has reached 1 billion users. OpenAI says ChatGPT has 1.2 billion weekly users.

The answer stands in for the visit, most of all on hard questions. In Pew Research Center's study of 68,879 searches, people clicked a result on 8% of visits when an AI summary appeared and on 15% when none did. They clicked a link inside the summary on 1% of visits. Long questions drew summaries far more often: 53% of searches of 10 words or more, against 8% of one- or two-word searches.

The questions a complex product raises are long ones.

Buyers rely on the answers without fully trusting them. Forrester reports that 94% of B2B buyers use AI in their buying process, and that buyers now name generative AI or conversational search as a more meaningful source of information than any other. Forrester also warns that answer engines "often deliver incomplete or unreliable information, creating mistrust," and that 19% of buyers using generative AI feel less confident in their purchasing decisions as a result.

New and specialized products are the ones AI explains worst. Language models are weakest on facts that appear rarely in what they learned from. Researchers found that models "struggle with less popular factual knowledge, and that scaling fails to appreciably improve memorization" of it (Mallen and colleagues, ACL 2023). Researchers at OpenAI and Georgia Tech showed that for arbitrary facts with no pattern to learn, a base model will hallucinate at least as often as the share of those facts it saw only once in training.

Google's own account of AI Overviews errors points to "data voids," topics with "a limited amount of high quality content." A new processor architecture or a category term coined last quarter is, almost by definition, a thinly documented fact.

The answer is becoming the place where people act. Google says its new checkout protocol already enables direct checkout for "hundreds of thousands of brands and retailers" across Google. In the same announcement, it reported that product attributes supplied by lululemon were used in 50% of relevant AI recommendations during testing: what a company tells the system shapes what the buyer is told. Forrester predicts that in 2026 at least one in five B2B sellers will have to respond to AI-powered buyer agents.

We have seen a shift like this before. In 1998, Sergey Brin and Larry Page argued that hand-built web directories "cannot cover all esoteric topics," and search replaced them; by 2008, Google was publishing its own Search Engine Optimization Starter Guide. A discipline grew up around being found. The new front door needs one built around being explained correctly: for a simple product, being found is most of the work, but for a complex one, the explanation is the work.

The world we are building toward

Last month I wrote that an AI-ready brand is "found, understood accurately, and chosen, by the engines and by the buyers they send." Understanding is the middle step: an engine has to find a product before it can explain it, and a buyer has to understand a product before choosing it. That middle step is the hardest to earn and the most valuable to own, and it is what SolCrys is built around.

When we succeed, four things will be true.

• Any answer engine can explain a complex product accurately, at the depth the asker needs. A CFO, an architect, and an engineer ask different questions about the same product. Each gets an answer that is correct, current, and in terms they can use, and each answer traces back to facts the company stands behind.
• Companies govern how they are explained. A company already governs its financial statements, its legal claims, and its brand guidelines, but not yet the explanation most buyers meet first. We expect that to change: a company's approved facts, claims, and limits become infrastructure, maintained like a codebase and used by every surface that explains the product.
• The buyer's question becomes the unit of marketing work. Not the campaign and not the keyword, but the question a buyer brings, and whether the answer to it is right. Progress is measured as a verified change in how the product is explained.
• Buyers decide with less fear. If understanding is the step between found and chosen, better explanations shorten the stall. Fewer good products lose to "no decision" because a buyer could not explain them to their own committee.

Answer engines are where we start, for two reasons. They are the newest explainer and the least governed one, and they are the one place where an explanation can be measured the same way, again and again, at scale. The discipline does not stop there: the same facts should explain the product on the company's website, at its booth, and in its demos, and to the software agents that will increasingly research and buy on a person's behalf.

The method: the Loop

Our method is a closed loop of four steps, Measure, Diagnose, Execute, and Verify, run again and again against the same buyer questions and the same approved facts.

Why a loop and not a campaign? Control engineers put the principle in one line: "base correcting actions on the difference between desired and actual performance" (Åström and Murray, Feedback Systems). A campaign is open loop: it publishes and moves on, without checking whether the explanation changed.

That fails for answer engines, because the explanation keeps moving. Models are updated, competitors publish, and sources age, so a fix that is never checked decays without anyone noticing. A closed loop is how an explanation stays correct.

• Measure. Build a set of buyer questions around who is asking, what they need, and how far along they are. Run them on ChatGPT, Gemini, Google AI Overviews, Perplexity, and Claude, and record how each engine explains the product today.
• Diagnose. Find where the explanation breaks: the product is absent, the right sources are not cited, the facts are wrong or out of date, or a competitor's framing wins. Answer Accuracy grades answers against the company's approved facts, and Workspace Deep Analysis looks across every question to find the root causes behind many symptoms.
• Execute. Turn the diagnosis into work grounded in Corporate Context and routed for human review: a clearer category page, a missing comparison, a corrected fact, a demo that shows what a paragraph cannot. The company's team publishes the work, or SolCrys builds it for them.
• Verify. Rerun the same questions against the baseline, judge the change across several runs, and feed what changed into the next round.

Two details make this a method rather than a slogan. First, the questions and the facts hold still within a turn, so before and after are compared on equal terms. Between turns, the facts grow as Diagnose surfaces what the company never wrote down.

Second, Verify is deliberately patient. The same textbook warns that feedback brings in "sensor noise" that needs "careful filtering," and answer engines are noisy sensors: the same question can get different answers on different runs. Judging across runs is how we avoid chasing noise.

W. Edwards Deming insisted that his improvement cycle be called Plan, Do, Study, Act, not Check. The spirit carries over: each turn is something we study and learn from, not a box we tick.

Cornelis Networks shows what this looks like with a deeply technical product. It needed buyers and engines to understand a brand-new networking category, so we told one story at three depths, from the homepage to the category page to a reference architecture, and measured each piece. "SolCrys AI took a deeply technical concept and turned it into real content mapped to different technical levels, and measured the effectiveness of every piece they produced," said Brandon Draeger, chief marketing officer of Cornelis Networks. At the company's trade show booth, an interactive demo turned the reference architecture into "something a visitor could grasp in two minutes," said Nishant Lodha, senior director of marketing (case study).

Why the work compounds

The Loop is a discipline, not a moat. Anyone can draw four verbs on a slide. What compounds is what each turn of the Loop leaves behind.

A governed source of truth. Corporate Context holds a company's approved facts, claims, and guardrails, set once for the whole organization. Each turn adds to it or corrects it, because a wrong answer found in Diagnose often exposes a fact the company never wrote down. Over time it becomes the company's most complete, current description of its product, and the one source for every output, whether the first reader is a person or a model.

Comparable history. Every action is tied to the same buyer questions, measured the same way before and after, so the record shows what changed after each piece of work, on which engine, for which kind of question. A company that has run the Loop for a year knows how its category gets explained in a way a newcomer cannot buy. Across many categories, we learn which kinds of work tend to move answers: patterns, never one customer's private facts.

A harness that outlives any model. The models we use will change many times. What carries over is the harness around them: the scenarios we test against, the labeled examples that define a good answer, and the record of how each model performed on our jobs. I made that argument at length in Models Change. The Harness Compounds.

Trust that is slow to build. Companies with complex or regulated products will not let an outside system say things about them without review. Approval, and a record of where each claim came from, have to be designed in from the start. They are hard to bolt on later, and they are what lets a cautious company say yes.

A path from diagnosis to the finished fix. Many teams can see the gap but lack the people to close it. When a fix needs a new category page, an interactive demo, or a booth animation, SolCrys Sites and SolCrys Motion build it from the same Corporate Context, so the explanation stays consistent wherever the buyer meets it.

For a customer, these assets grow more valuable with every turn: a longer verified history and a fuller set of governed facts. For SolCrys, each customer and each category make the method better for the next one.

What has to be true

A vision is a set of bets. These are ours, with the evidence we have today and the signal that would tell us we are wrong.

• Answer engines stay a primary explainer for considered purchases. Evidence today: Weekly AI use for information rose to 38% in 2026, and 94% of B2B buyers use AI while buying. We are wrong if buyers of complex products move their research back to vendor sites, analysts, and people.
• What a company publishes still shapes how it is explained. Evidence today: Google's account of AI Overviews errors points to "data voids"; brand-supplied attributes were used in 50% of relevant recommendations in Google's test; and our customers' category questions moved after they published. We are wrong if engines lean so heavily on third-party, community, or paid sources that a company's own facts stop mattering.
• Companies pay for verified change, not for dashboards. Evidence today: Basic tracking is already offered free by many tools, so measurement alone is not where the value sits. We are wrong if teams buy cheap tracking and do the rest by hand, without losing ground.
• Accuracy and governance beat raw speed. Evidence today: Buyers report mistrust of unreliable answers, and technical and regulated companies require review. We are wrong if engines reward volume over accuracy, and fast, ungoverned content wins for long enough to matter.
• Understanding can be measured, not just asserted. Evidence today: We can grade how engines explain a product against approved facts. We are wrong if we cannot connect better explanations to better buyer decisions.

The last bet is the one we have proven least, so I want to be plain about it. Today we measure how answer engines explain a product: whether it is mentioned, which sources are cited, and whether the claims match the company's approved facts. We do not yet measure what a buyer understands after reading the answer, and building that measure is part of the work ahead.

The early results point the right way. Working with Cornelis Networks, we helped make a brand-new networking category answerable in four weeks: the mention rate on that category's questions rose 12-fold, on a term nobody had published before, and the company's overall AI mention rate rose nearly 7 percentage points (case study). NextSilicon's AI mention rate on HPC and AI questions rose from 1.9% to 7.4% in 45 days (case study).

These are early results from two categories. As the evidence grows, we will publish it with its sample sizes, by the standard I set last month: a lift shown without a sample size is a picture, not a finding.

Questions we hear

Isn't this SEO under a new name?

SEO earned its place by helping companies get found. For a complex product, the harder problem comes next: being explained correctly, and that now happens inside the answer. The work overlaps with good SEO, such as clear, well-structured pages, but the unit is a buyer's question and the measure is the answer itself.

Won't the answer engines fix this on their own, or take facts straight from brands?

An engine can only explain what is written down somewhere it trusts, and the facts about a new product have to come from the company that makes it. As engines build ways to take those facts in directly, the question shifts from whether a company supplies facts to which facts it supplies, and whether they are approved, current, and consistent. That makes a governed source of truth more valuable, not less.

How large is this?

Every company with a complex product already pays to be explained: its website, product marketing, content, events, analyst relations, and sales engineers. What is new is that the most influential explainer is a system none of those budgets governs. We expect governing it to become a standing part of how those companies spend.

Why SolCrys?

We built for the hardest case first. Our deepest work is with products that are hard to explain, in high-performance computing and AI infrastructure, where a vague answer costs the most. Governance was there from the start: Corporate Context holds each customer's approved facts, and Answer Accuracy grades engine answers against them. Many tools now track whether engines mention a brand; our bet is on what comes after tracking: governed facts, finished fixes, and verified change.

How we work, and what we refuse

A vision only matters if it changes what we do on an ordinary Monday. Alignment is rarer than it looks: in a study of 124 companies, only 28% of executives and middle managers responsible for executing strategy could list three of their company's strategic priorities. Ours fit in one line: make complex products understandable, run the Loop honestly, and let what compounds compound.

These principles are how we decide.

• Accuracy before reach. We never make a product sound better than it is. A buyer who understands a product correctly and walks away is a better outcome than one who buys on a misunderstanding.
• Start from the buyer's question. We work backward from what a buyer needs to understand at their stage, as Amazon's principle puts it: "start with the customer and work backwards."
• Claim only what we can verify. Same questions, same baseline, judged across runs, with the method and the sample size stated.
• Facts are governed, and people approve. Anything we draft for a customer is grounded in their Corporate Context and goes to their team for review before it ships.
• Choose models by evidence, not reputation. We compare models on our own scenarios and labeled examples, and let the results guide the choice.
• The process is not the thing. Jeff Bezos warned that a process can become "the proxy for the result you want." Running the Loop is not the goal. A buyer who understands the product is.

And these are the things we will not do, even when they would move a number:

• Manipulate answer engines with hidden text, planted instructions, fake reviews, or coordinated posting. Researchers have shown that a "strategic text sequence" can push a product up an AI's recommendations, and that is exactly the kind of win we refuse.
• Invent or inflate a claim, a rating, or a number, for a customer or for ourselves.
• Publish anything under a customer's name without their review.
• Call a single run's movement a result before we have rerun it and judged it across runs.
• Attack competitors with claims we cannot source.

We will debate these bets hard. Once we decide, we disagree and commit. Whatever their role, everyone at SolCrys should be able to name the step of the Loop their work serves and the evidence that will show it worked.

The goal is simple to state. When a buyer asks an answer engine about a complex product, the answer should be one its makers would sign.