The Need for Speed. (AKA – Your commerce search is broken)

Most merchants still can’t tell you why they should move to AI search. Not because the case is weak. Because nobody has made it plainly enough.

The why still isn’t clear.

Ask ten merchants why they’d swap keyword search for AI search and you’ll get ten fuzzy answers. “It’s more modern.” “Customers expect it now.” “Our competitors are doing something with AI.” All directionally true, all too vague to justify a budget line.

Here’s the clear version. AI search wins because it is faster to a correct answer, and speed to a correct answer is the single variable that most determines whether a shopper buys from you or bounces to someone who answered faster. Everything else, natural language understanding, exact color matching, image similarity, is in service of that one outcome: getting the shopper from intent to product in the shortest possible time. We’ll give that outcome a name shortly. First, the full list of why AI search gets there faster than the two things it’s replacing.

Every advantage, spelled out

Trimodal search, natural language, precise color, and image similarity, beats the two systems it replaces for specific, listable reasons. Not vibes. Reasons.

Versus keyword search

Understands intent, not just string matches. A keyword box matches characters. Natural language search matches meaning, so “something for a rooftop wedding, not too formal” returns real results instead of zero.

Accepts full sentences, not query fragments. Shoppers don’t think in two-word queries. AI search lets them type the way they’d describe it to a friend, which is how they’re already searching everywhere else.

Removes the vocabulary requirement. A shopper doesn’t need to know your taxonomy’s word for it. They don’t need “midi” versus “maxi,” “trench” versus “overcoat.” They just need to describe the thing.

Understands compound, multi-attribute requests in one pass. Color, fabric, occasion, and silhouette in a single sentence, instead of one keyword, one filter, one attempt at a time.

Resolves subjective and emotional language. “Vibe,” “mood,” “old Hollywood,” “not too casual but not black tie either.” Keyword indexes have no field for that. AI search does.

Matches color as data, not as a label. “Forest green,” “sage,” “olive,” and “dark green” are four different words fighting over one bucket of pixels. Precise color search matches the actual shade, not whoever tagged the product that day.

Searches with no words at all. Image similarity search doesn’t just outperform keyword search, it does something keyword search structurally cannot do. When the shopper has a photo and no vocabulary, keyword search is not a worse option. It is not an option.

Cuts the zero-result dead end. Zero-result rates on keyword search average 10 to 15% industry-wide. Every one of those is a shopper who typed something real and got nothing back.

Versus hierarchical navigation

Removes the burden of guessing your taxonomy. Faceted nav only works if the shopper correctly guesses which department, category, and subcategory you filed the product under. AI search skips the guess entirely.

Collapses a multi-click decision tree into one interaction. Department, then category, then subcategory, then five filters, then sort, is a lot of decisions before the shopper sees a single product. One description replaces the whole tree.

Eliminates filter fatigue. Every checkbox, every slider, every chip is a small tax on patience, and every one triggers a page reload. AI search charges that tax once.

Handles intent your taxonomy was never built for. “Something for my nephew’s graduation” doesn’t map to a category. It spans several. Faceted nav siloes; natural language search doesn’t care about your site map.

Avoids the over-filtered dead end. Combine enough facets on a traditional nav and you land on an empty grid, no matter how deep the catalog is. AI search resolves ambiguity instead of returning nothing.

Removes the backtrack tax. Pick the wrong facet path and the only way out is back, then re-filter, then hope. Get an AI search result wrong and the fix is one more sentence.

Every one of those is a real, structural advantage, not a nice-to-have polish. And every one of them saves time. That’s not a side effect. That’s the point.

ChatGPT already settled the argument

If there’s still a doubt in the room about whether AI-driven search beats the old model, ChatGPT closed it. It now fields on the order of 50 million shopping-related queries a day, roughly 2% of its total prompt volume, and that share has climbed steadily as shoppers learned it simply works better than typing into a search bar. Nearly 78% of consumers say they’ve used AI for shopping in just the past six months. Close to 60% now use generative AI for shopping specifically, and over 40% do it weekly or more.

That shift shows up in the money, not just the usage. AI-referred traffic to U.S. retail sites is up 138% year over year as of this spring, and up more than 1,300% since late 2024. And the traffic isn’t just bigger, it’s better: AI-referred shoppers now convert 54% better than shoppers from non-AI sources, and they spend 53% more time on site browsing 23% more pages once they arrive. A year earlier, that comparison ran the other way. It flipped because the search got better, fast.

Without a shadow of a doubt, AI queries handle search better than what merchants are running today. The only open question is how fast each merchant closes the gap.

Give the metric a name

Time-to-Discovery, TTD – The interval between a shopper arriving with intent and that intent resolving into a product they can actually buy.

Every advantage above is really a TTD advantage in disguise, fewer dead ends, fewer clicks, fewer reformulated queries, fewer page reloads. Relevance still matters. But relevance that arrives too slowly loses to a decent answer that arrives instantly, because the shopper with options rarely waits around to find out if the slow answer would have been better.

What “successful” used to cost

“Successful search” isn’t one moment, it’s at least three, and each one has always cost more time than merchants like to admit.

Click-through to a product page. This is the cheapest win, and traditional search still fumbles it constantly. Zero-result rates average 10 to 15%, and 76% of U.S. shoppers say they’ve hit an unsuccessful search on a retail site, averaging four failed searches a month. Only 12% of shoppers say they get exactly what they’re looking for every time. Every one of those failures means typing again, reloading again, waiting again, each reload bounded by a patience clock that’s already thin: 40% of shoppers won’t wait more than three seconds for a page to load, and bounce rates jump from 9% under two seconds to 38% at five. Historically, reaching a product page wasn’t a one-shot, sub-second event. It was an iteration loop, type, scan, refine, reload, repeat, and every lap cost real seconds shoppers didn’t have to spare.

Add to cart. When search actually works, the payoff is immediate: 92% of shoppers who complete a successful search buy the item they were looking for, and 78% add at least one more item on top of it, three more on average. Searchers convert at roughly 1.8 to 3 times the rate of shoppers who only browse. The gap between a shopper who finds the product and one who doesn’t isn’t a small optimization. It’s the difference between a sale and an empty cart.

Purchase. This is where the historical cost compounds hardest. Failed on-site search is estimated to cost U.S. retailers on the order of $234 billion a year. More than half of shoppers who can’t find an item they want abandon the cart and go elsewhere entirely. Eighty-one percent leave the site after an unsuccessful search, and 82% say they avoid returning to a site where search has failed them before. Every one of those numbers is a shopper who had money to spend and ran out of patience before they could spend it.

2 B.C. vs. now

Call it 2 B.C., two years Before Chat. Shopper patience for slow, clunky discovery hasn’t just declined since then, it’s been recalibrated entirely. The habits that used to define “good enough,” a keyword box, a filter panel, a bit of scrolling, were formed in a world where nobody had a faster alternative sitting open in another tab. That world is gone. AI-referred retail traffic is up over 1,300% since late 2024 alone, and the shoppers arriving through it now convert better than shoppers from any other channel, a complete reversal from where that comparison stood a year earlier. Shoppers didn’t get more patient with your search bar. They got a faster one everywhere else, and they now measure your store against it whether you’ve upgraded or not.

AI search TTD, by the numbers

Here’s what a low TTD actually looks like in production: four search types, four real results, each returned in seconds, not minutes. Results from visualAI’s trimodal search engine.

Demo video

Precise color search – Roughly 1 to 2 seconds from color selection to a full grid of closest-shade matches. No guessing whether “bright green” means what the merchant meant by it, the shopper picks the color shade and the engine returns it.

Natural language search – Roughly 1 to 3 seconds from a full-sentence description to relevant results. “I’m after a floral dress made of lightweight materials like cotton or linen, brightly colored, midi length, short sleeves” returns exactly that, not a keyword-matched approximation of it.

Natural language plus color – Last week in this space we called color the next frontier in agentic commerce, specifically the moment an exact shade and a plain-language description resolve as a single query instead of two separate steps. That combination is a TTD story too. A shopper isn’t forced to pick between describing what they want and specifying the color they want, they do both at once. “Mini dress” paired with a dialed-in red still returns matched results in roughly 1 to 2 seconds, the same range as either mode alone, because resolving two signals together is still one query, not two round trips.

Image similarity search – Roughly 3 to 4 seconds from an uploaded photo to visually similar in-stock products, the heaviest lift of the three, and still under four seconds. A shopper with a picture and zero product vocabulary gets a full page of real, buyable matches.

Of course, all of the above results are subject to catalog product depth. But every catalog, regardless of size, can benefit by adding AI search. 

That’s the whole customer experience upgrade in one sentence: a shopper who used to spend minutes iterating through failed queries and filter panels now gets a correct answer before they’ve finished reading the results page. Lower TTD doesn’t just feel better. It shows up directly in the numbers above, more click-throughs, more add-to-carts, more completed purchases, because the shopper never had the chance to lose patience and leave.

The 1-2% and the 28M

The major marketplaces and the handful of super merchants get all of this. They’ve already shipped natural language search, and a growing number have added image similarity on top of it, because they can see the same TTD math everyone else can see. But that group represents something like 1 to 2% of the commerce market. The large tool vendors building this technology are occupied serving that same sliver, the Fortune 500 accounts with the budgets to match. Nobody at that scale is building a mass-market, cost-effective version of this for main street, the roughly 28 million online commerce destinations that make up the rest of the market. That gap is exactly why we built discoverGPT.

Two weeks ago, in this space, we said it plainly: it is not all about agentic readiness. Agentic data and feed preparation are table stakes for off-site discovery, the work that makes a catalog legible to ChatGPT, Gemini, and the marketplaces. But off-site discovery is a needle-in-a-haystack game, a single merchant competing against billions of listings for a slot in someone else’s ranking. Onsite is the one place a shopper’s query is guaranteed to surface one of your own products, and it is core, critical, and unaddressed on the vast majority of the 28 million stores that make up main street commerce. Every day that discovery stays slow onsite is another day of daily visitors quietly handed to agents, marketplaces, and giants who already fixed their TTD.