AI-first is Hard 

At the UXforAI blog, we have been covering AI-first design patterns since the beginning. Starting with a simple chat using the first iteration of ChatGPT (https://www.uxforai.com/p/the-art-of-ai-conversation-ai-chat-pattern) and then again with the r1 rabbit (https://www.uxforai.com/p/the-rise-of-ai-first-products) and then again with “Apps Are Dead” (https://www.uxforai.com/p/apps-are-dead) 

Much of what can be considered “AI-first” today is based on chat, and as Jakob Nielsen so rightly pointed out, “chat is the new command line.” (https://www.nngroup.com/articles/ai-paradigm/) While the NLP command line is undoubtedly much friendlier than the Linux command line, it does not make for a complete application for most targeted use cases. Even with such chat-forward applications as “AI Girlfriend,” you still need to occasionally set up your new virtual Tamagotchi, add outfits, take her to dinner, and whatever other maintenance and editing tasks such a virtual companion requires… Definitely subscription payments and such… (I wouldn’t know… No, honestly! Those vile rumors are widely exaggerated! I swear!)

Ahem, the larger point is that a chat alone is not a complete interface for most use cases, no matter how we slice this. Neither is a Copilot. As we discussed at length (https://www.uxforai.com/p/ux-best-practices-copilot-design), Copilot, as it is most frequently implemented in a sidebar, is not an AI-first application. It is, at best, an add-on to the existing pages meant to serve in an advisory and helpful capacity while the Information Architecture of the original app is preserved. 

All these limitations of chat are supposed to be now transcended with the pattern du jour: the Canvas.

Design Pattern du Jour: The Canvas

The Canvas (https://openai.com/index/introducing-canvas/) introduced by ChatGPT is all the rage, with many designers seeming to think it will solve all their present and future information architecture problems:

Again, I’m sorry to disappoint you – just because you can now pop your virtual girlfriend out of the chat in a new Canvas, will not remove the need for a girlfriend personality edit screen, welcome experience, subscription level screen, payment history, etc. 

As a general rule,

It is essential for UX designers in this age of daily AI novelties not to copy blindly but instead to take the time to understand the capabilities of the new AI models and get creative by putting them to use for specific use cases that matter to your customers. 

(And if you need inspiration from the great inventors who changed the course of history, check out https://www.uxforai.com/p/ux-for-ai-guide-to-inventions-and-patents-part-1)

Is Information Architecture Dead? 

In the age of AI chat’s proliferation, many UX folks seem to think that the Information Architecture is dead. I’d like to introduce a different idea. Based on my research with 34 AI-driven projects, if your AI-first experience starts and ends with chat, customers often have trouble understanding what your application does and the specific value it can deliver. Customers also need help coming up with queries that deliver the most value. Depending on your application, customers also need to be made aware of certain time-sensitive developments, like alerts, sales, etc. 

Chat does not apply equally to all cases because it often simply waits for the user to ask a question, which is far from ideal for every single application because: 

And while modern LLMs are so much better at interpreting imprecise instructions than any computers so far in the history of humanity, sometimes you want a pre-determined set of queries and decent starting points even to understand what is happening in your app. This is especially important in the digital realm, where we do not have access to environmental context and have no shared history with the AI agent. 

In other words, 

Fortunately, this problem can be reasonably easy to solve with my new AI-first Information Architecture framework. However, I should warn you up-front (and repeat this ad-nauseam): 

To demonstrate this new approach, let’s compare the Information Architecture for the Conventional Amazon pages with what might be imagined as AI-first Amazon.

Amazon – Conventional Approach

Let’s use the quintessential example of an e-commerce site, going back to one of the very first examples I had ever written about in my very first book, Designing Search: UX Strategies for eCommerce Success (https://a.co/d/8w6s6F9) back in 2011 – the mighty Amazon. 

For the purposes of this exercise, we can divide the site roughly into four parts: Homepage, Search Results Page, Item Detail, and Maintenance (which includes things like Past Orders, Checkout, Returns, Payments, Addresses, etc.)

“AI-minus”? Homepage

As you are no doubt aware, Amazon is already largely an AI-first application, and the homepage easily reflects this, tempting the users with sales based on stuff the algorithm thinks they might like:

Unfortunately, the algorithm seems to have no clue about what I actually need or want, as we discussed at length here: https://www.uxforai.com/p/ai-search-part-2-eye-meat-and-doi-sort-algorithms 

For example, I’ve already purchased some lawn furniture, so how likely am I to buy more? How many tofu presses does one person need? And why does it stubbornly keep suggesting that I need more V for Vendetta posters? (Does it know that I, too, keep loads of illicit ultra-liberal (vegan?) butter, fine art, and Natalie Portman in my basement in preparation for the next four years of the new regime?)

Conventional Search Results Page

Search pages are the heart of a typical e-commerce site, frequently accounting for the lion’s share of the overall traffic. While Amazon’s results are the envy of the industry, the search largely fails for complex NLP queries such as “Mysteries that are not scary” (Jared Spool’s idea), as we have written about at length here: https://www.uxforai.com/p/search-ux-revolution-llm-ai-in-search-uis 

Just as last time we reviewed this query, today’s results also include a hodge-podge of hits. From children’s books to Stephen King’s very scary “If It Bleeds” to the presumably terrifying “Don’t Close Your Eyes,” which promises to be “An absolutely gripping psychological thriller with a twist that will keep you up all night.” 

Results that are almost, but entirely completely unlike what we actually asked for. 

Clearly, not many of the modern LLM capabilities are in evidence!

AI-plus Item Detail Page

In contrast to search results, the Item Detail screen is where the LLM capabilities of Amazon really come out to play:

AI drives much of this content, from “looking for specific info” question ideas to brand-specific trends and finally to a highly used and well-liked reviews summarization feature, which is admittedly excellent. Note that the main search box at the top of the page still retains our original query. If we wanted to search for something in this veritable monster of the page (it scrolls for over 20,000 vertical pixels), we’d have to use the browser page search (Command-F), the specific Q&A search box, or search user reviews – all separate search boxes!

Conventional Maintenance Pages

Finally, we also have a set of maintenance pages that today are completely tactical, with little or no AI features. These pages mainly consist of straight-up lists with time-based LIFO (Last-in-First-out) sort algorithms. Each of these pages also comes with its own targeted search box.

Now, let’s examine how a site like this might be redesigned to be AI-first using our new framework.

AI-first Amazon Redesign

As we discussed earlier in the article, simply enabling an LLM-based search, while a fine start, will not fulfill every single use case covered by Amazon. We can do much more! Nearly every feature of the new AI-first e-commerce site (let’s call it “AI-first Amazon”) could be driven by AI. 

So far, there is not much difference between the old and the new IA, but the devil, as they say, is in the details (and also in Georgia, AZ, FL, MI, NC, NV, PA, and WI… If you believe the recent election results. And we believe. Well, some of us do. This time. Maybe.)

Let’s take a closer look.

AI-first Analysis Overview Page

The old homepage can be replaced by an AI-first analysis page. While this may seem like a small shift, the change can be extensive. For example, this week in the US, it’s Black Friday – which means that all of the retailers make a concerted once-a-year effort to distract the populace from the usual right-wing grievances and massive hurricanes driven by global warming through the frenzied acquisition of discounted (but soon to be heavily tariffed) goods from China. 

Far it be for me to interfere with such a sacred American yearly tradition! In fact, I’d recommend doubling down on driving the usual frenzy using an LLM. Here’s a quick wireframe of what that might look like, courtesy of ChatGPT 4o – Just imagine some item image details (and maybe a dynamic shot of the crowd fighting to the death over the last Tickle Me Elmo), and you are there:

While the AI-first Analysis page might look very similar to the current Amazon page, there are some significant differences: 

• LLM text summary: mentioning the unique nature of this specific Black Friday – Tariffs, etc. 

• Why Shop Now: Reasons to shop that play on people’s current fears… (Because fear sells! Nice. Very apropos.)

• Gift/Discount Categories: specifically targeted at popular gifts and discounts, not at things the user might be directly interested in.

Impressive, no? This is literally what is accessible today using the non-fine-tuned LLM tech. 

But wait, there is more. Oh yes. So. Much. More.

AI-first Category Analysis Pages

With LLMs running, the system can now understand and detect what categories, places, things, or ideas the person is interested in. Try this little experiment: simply dump your past Amazon orders into ChatGPT and ask what categories and pastimes this customer is interested in and what they might want to buy next. I must admit that even with my reasonably inflated expectations, the modern LLMs have blown me away. ChatGPT understood not only the categories (fishing, gardening, working out, dog, etc.) but also the item specifics for purchases that logically followed. I guess I should not be surprised by this —

For example, since I recently bought a new light-weight spinning reel, ChatGPT determined that I now have two of them and I might likely need another rod, light-weight line, and trout lures, and came up with excellent recommendations for each, without any help from external sources:

Some things to note about this page:

• Category-Item-Specific LLM Summary: the summary is again on the top right of the page. However, this particular summary is much more specifically targeted to my recent shopping and interests, as is expected of the category page. 

• Side Image: so much more than just visual interest, this image is generated by Midjourney as a dynamic prompt from ChatGPT based on the page’s content. It is a custom smorgasbord featuring some of the beautiful products that are mentioned in the body of the page. (Yeah, it’s a bit rough, but that will get better fast.)

• Order of the Items: again, that is self-selected by ChatGPT to reflect what you are most likely to need right now. 

Now let me reiterate: 

Isn’t that just a little better than a bunch of items? You bet. It’s telling you about the adventure of landing the fish of your dreams using the items, not just the items themselves. (NOTE: you also will have the individual images, of course – this is just the hero image tying them together.)

This is the equivalent of a visual crack cocaine for a fisherman. And it took me just minutes to “design” (more like “put together” or “orchestrate”) based on the ChatGPT’s output. Now imagine each and every category you really care about meticulously analyzed, likely items on sale highlighted, with a custom smorgasbord of items specially selected for your interests and shopping habits. The kind of personalization knocks the current experience completely, ahem, out of the water. 

Right then, moving on.

AI-first LLM Search

We have already covered this earlier in the series, but it bears repeating. LLM can easily solve for NLP queries that require complex reasoning, such as “Mysteries that are not Scary” (thanks again, Jared Spool, it’s a gift that keeps on giving):

Just imagine adding the item images and you are there.

Here are a few things to note about this page:

• Machine Interpretation of the query: “Enjoy these intriguing tales that promise suspenseful fun without sleepless nights!” (at the top of the page). It is playful and precise. It captures the spirit of what you are asking for in a natural language. No e-commerce site does this today (yet). We covered re-stating patterns here: https://www.uxforai.com/p/llm-design-patterns-re-stating-auto-complete-talk-back-suggestions-nest-steps-regen-tweaks-and-guard 

• Result-Specific LLM Summary: this summary is even more remarkable because it’s essentially constructed on the fly using the interpretation. 

• Exceptional Results: check.

• “Scariness rating”: a custom metric that, together with the item summary (explicitly created to help the user interpret the results of their NLP query), helps the user evaluate the item’s fitness of purpose and zero in on the exact item they did not even know they wanted. 

• Facets preserved: no reason to drop the old baby on its head just because it’s mixed together with the tepid bath water of conventional search results. Facets are highly beneficial and are frequently used by many searchers. There is no reason they need to be omitted because we are using the LLM – in fact, the facets themselves can be augmented dynamically by the LLM to include the said “Scariness rating” and other custom metrics, as appropriate. They can include a much richer subset of facets dependent not only on the item metadata but also on the query itself – a brand-new search instrumentation capability!

• Custom Imagery: again, here we have a Midjourney-generated custom image built using a summary prompt from ChatGPT.

In short, these LLM-produced results are not just superior in content (with no input from me whatsoever) but also superior in experience. This AI-first search results page makes an effort to tell the story – in the context of the customer’s question. AI is talking with the customer using their language. It’s not just accessible; it’s individualized (See this brilliant blog by Jakob Nielsen:  https://jakobnielsenphd.substack.com/p/accessibility-generative-ui). AI offers explanations, ratings, and filters appropriate to the task it understands the customer is trying to solve.

Task-related queries that require multi-level conversations and specific refinements, such as “What items will I be likely needing for my Patagonia trip?” or “I need a door gasket that fits my classic Jaguar,” should also be no problem. Welcome to the next level, baby!

AI-first Item Detail with Contextualized Search

Forget the conventional item details, where each section (Q&A, Details, Reviews) needs to be searched using its own search box. In the new AI-first paradigm, we can create a simple all-in-one search and Q&A experience using the “ask bar” on top of the screen that has been contextualized to this specific item using an editable/removable orange tag in place of the usual category drop-down:

Things to note about this page:

• Contextualized, all-in-one Q&A and search: Forget the ye olde paradigm of multiple search boxes! Now that the search/ask has been contextualized to this item, the top bar does everything you need, including answering questions about item specifics and speculating (or even visually demonstrating!) how the item would fit into your active lifestyle. (Need to search for something else? Just delete the orange tag. Want to ask a question about a different item? Click on the orange tag to change it.)

• Item-Specific, User-Specific LLM Summary: this page also has a text summary, but this one is specific to the item and also ideally created by anticipating the questions this specific user is likely to want to ask. Does your user have wide feet, pronation, or plantar fasciitis? The summary will talk about the item as it relates to the user’s concerns.

• User-Specific next steps questions: note the blue tags below the summary – these are additional questions not included in the summary that are a combination of what the user might want to ask next and also what other users similar to him/her have asked next after seeing the summary, plus the usual DOI “AI smarts.” It’s a powerful combination that will form the “secret sauce” of the new competitive landscape. (Yes, UX should help formulate this: https://www.uxforai.com/p/ai-search-part-2-eye-meat-and-doi-sort-algorithms) 

• User-Specific imagery: Maybe feature the item together with things the user purchased previously? (Not this image – I don’t buy white clothes. EVER. Period. White watch? Seriously? Ugh.)

• User-interest-specific imagery: Maybe show a slightly geeky (but good-looking!) clean-shaven guy doing some fishing, wearing the new shoes? (Again, NOT this specific image – I do not run on trails. I hike. It’s much more dignified.)

The point is that this summary again tells a compelling story about how the item relates to the user – this is what it means to be AI-first. 

NOTE: This could also cause some very silly and mistaken combinations. For example, if a wife buys a lot of power tools for her husband. On the other hand, this could be quite entertaining – users might let AI keep hallucinating weird activities just for entertainment! OK, OK, we’re kidding! Actually, there already is a simple and well-accepted way to solve this problem: simply avoid all the AI-related guessing nonsense and allow the user herself to pick the profile of who they are shopping for, just as Netflix or Amazon Prime does with movie picks. Because not everything needs to be AI:

AI-first Q&A Maintenance Pages

The magic of LLMs can also apply to maintenance pages. For instance, answering order-related questions such as: 

"When will my order arrive?"

"What does 'Processing' or 'Shipped' mean?"

"Can I change or cancel my order?"

"How do I return an item?"

"When will I receive my refund?"

And many more questions are better asked in a normal, conversational Q&A format, the new LLMs’ bread and butter. By now, you can probably fill in the rest of the features: contextualized search, an LLM summary relating to orders, a mini-report full of ideas about shopping for necessities in bulk and saving money (while, of course, spending more… Much, much, MUCH more! BWAHAHAHA!)

IA is Dead. Long Live IA!

In conclusion, everything we’ve covered today is directly accessible right now. All it requires is someone to dream up these new experiences and come up with simple wireframes. (Well, that’s the easy part, actually – convincing the leadership to prioritize these new types of experiences and invest in the tech stack that will make them truly sing… While discarding the conventional designs – now that will be the hard part. Good luck.) However, we hope that today we helped you a little with the basics of Information Architecture that can be considered AI-first. 

Take the time to understand the importance of contextual storytelling, identification of categories, concepts, and ideas, contextual summaries, federated conversational search, constrained search, and next steps – all that makes this Information Architecture framework work as a new AI-first UX play.

Most importantly, try to see how this new framework delivers a superior experience:

None of the new techniques we discussed here mean discarding the wisdom and insight collected in the Information Architecture of conventional applications — but they do involve thinking about those conventional insights in a new AI-first way. Hopefully, our new IA framework will help you out at work, while whatever small entertainment value we provide helps you carry your load just a little lighter this Holiday season. 

Drop us a line and let us know what you think.

Long Live Information Architecture!

Greg Nudelman (with Daria Kempka, Contributing Editor)