Selva
Selva is the decision layer for commerce. A decision model reads what your shopper means, code decides the answer from your own source of truth, and it asks one question only when one fact would change the answer.
Example stores and products are invented. A real store goes live only after Selva passes a test written for it.
The film
See how it decides
Example stores and products are invented. A real store goes live only after Selva passes a test written for it.
Under the bar
Missionswhat a shopper came to do
Benchmarksevery place an answer can go wrong
Casesevery answer written before the run
387 cases, 5 runs each, no wrong answer, median answer in 119 ms. Selva’s internal calibration on Northfork, its invented auto-parts store, November 11, 2026.
Your store gets its own: the missions its shoppers come to do, and the benchmarks every answer has to pass, written for your category.
A question it can’t settle never stays that way. It becomes a test, then a capability, in every store.
Worked examples
It sits between your catalog and your shopper. It reads the whole request, checks it by code against the maker’s own fitment documentyour own size chart, order-by windows and policy, and gives the same verdict to search, the product page, the cart and any AI agent that shops your store. Your platform keeps the records.
Selva is new, and opening to its first stores. Every figure on this page about Selva comes from its internal simulation or its internal calibration until a store measures it, and says so.
01Reads the whole sentence
Shoppers now type the whole job, not a keyword. Google says the average search in its AI Mode is triple the length of a traditional query. Keyword search keeps the partthe dress and drops the rest: the underlined words, which are usually what decides the sale.
Google, May 19, 2026 (source). The requests below are examples.
One understanding
Search
Search
Chat
Will this cover fit my 2022 Ram 1500?
Which bed, the 5′7″ or the 6′4″?
5′7″
It fits. I left out the 6′4″ cover and liner, because bed length has to match.
Chat
Will this dress fit? I’m usually a 6.
What’s your bust, with a soft tape?
37 inches.
A 10. Your bust decides it, and the waist can be taken in. I left out the 6 and the 8.
Product page
Fits your 2022 Ram 1500, 5′7″ bed
Product page
A 10 by your bust, ready for June 12 if ordered by March 13
Sales desk
R-1182The same answer the shopper saw, word for word.
Sales desk
10The same answer the shopper saw, word for word.
Assistant feeds
Assistant feeds
How it decides
The shopper asks
“Will this cover fit my 2022 Ram 1500?”
Vehicle2022 Ram 1500Partfolding bed cover
Code checks the source
R-1182It asks once
Which bed, the 5′7″ or the 6′4″?5′7″
The verdict, everywhere
Fits the 5′7″ bed. The 6′4″ cover is left out.
SearchProduct pageChatCartSales deskAgent feeds
The shopper asks
“Will this dress fit? I’m usually a 6.”
Wearerusually a 6Itemchiffon bridesmaid dress
Code checks the source
10It asks once
What’s your bust, with a soft tape?37 inches.
The verdict, everywhere
A 10. The 6 and the 8 are left out.
SearchProduct pageChatCartSales deskAgent feeds
02The same answer, for AI agents
Adobe counts visits sent by AI assistants to U.S. retail sites up 62% in the year to July 2026. Those visits convert 60% better than other traffic. More and more, the assistant does the shopping too. It reads your pages faster than any shopper, and believes them more. If a page says “fits” at year, make and model, the agent buys on it.If the page doesn’t say which measurement decides the size, the agent orders the size it was told. With Selva, it gets your store’s own answer as data instead, so your customer is done in one exchange, not a guess and a return.
Adobe Analytics, July 2026 (source).
“4.10 gears for a 2019 F-150 5.0, going to 35s”
“A sage bridesmaid dress, size 14, for a wedding on June 12”
Your store today
What your customer reads“Both gear sets fit a 2019 F-150. The 8.8-inch set is $40 cheaper.”
Wrong axle shipped. A return that takes 9 to 10 days, and the install kit never sold.
With Selva
What your customer reads“Twelve bolts means the 9.75-inch set. It fits, from the maker’s own guide, and the job also needs the install kit.”
Right gears, and the kit. $668, kept.
Your store today
What your customer reads“Done: the sage dress in size 14, $128.”
The 14 won’t close at the bust. The exchange is a new made-to-order dress, and the wedding comes first.
With Selva
What your customer reads“At 41 inches, this store’s chart makes it a 16. Ordered today, it’s ready for June 12.”
Right dress, right size, on time. $128, kept, and the party stays whole.
Passes. All three runs are right.Fails. A price, stock or date your store doesn’t state sinks the whole question.A price or stock your store doesn’t state, or a date its own times don’t give, sinks the whole question.
Fails a run
Still passes
It doesn’t, directly. B21 runs on Selva’s own test store, through its agent tools, and shows what an agent gets when the answers come from Selva. Your own channels are counted separately: your feed against your pages, and what assistants say about your store when a person asks them in their own apps.
As exposure, not yet counted: your AI-referred sessions, times the share of questions agents get wrong, times your conversion and average order. Your AI-referred share comes from your own analytics. Selva scores this benchmark through its own agent tools, and never asks a public assistant about your store.
This page is written for agents too: runselva.com/llms.txt
03Counts what your reports can’t see
Selling is half the job. The other half is knowing what is happening in your store. Your reports count what comes back: returns, tickets, searches with no results. Some of the costliest fit losses send nothing back at all.
Fit losses per 1,000 sessions at a store that says “fits” at year, make and model.What wrong answers cost a typical fashion store, per 1,000 sessions.
What your reports see · $50Fit returns. Each one comes back, with a label and a reason code.Returns, traced · $50Each one tied to exactly what the shopper was told before buying.
What they never see · $92Hidden fits: the maker lists the part, your store says no, and the shopper leaves. No return, no ticket, no review.Hidden fits, counted · $92Every time your store hides a part the maker lists for that vehicle, counted per 1,000 sessions and priced in your margin.
What your reports see · $47Returns and contacts. Each one comes back, with a label and a reason code.Returns, traced · $47Each one tied to exactly what the shopper was told before buying.
What they never see · $118Orders that never happen: the shopper asked something the page couldn’t settle. Nine in ten leave without a return, a ticket or a review.Lost orders, counted · $118Every shopper who leaves at a question your page couldn’t settle, counted per 1,000 sessions and priced in your margin.
Counting it is the report’s job. Winning the sale back is the answer’s job, measured in the proof.
From Selva’s internal simulation of 300,000 shopper sessions, in its benchmark library of November 3, 2026, at the floor, where fit returns are anchored at 4% of orders; there, hidden fits are the largest single loss. At the unanchored ceiling, wrong-part returns are larger. Hidden fits rest on one estimate, 8% of maker-listed fits hidden, which the free check measures on your own catalog from your public pages.
From Selva’s internal simulation of a typical apparel store: a $125 order, 35% margin, $18 per wrong cart, $7.20 per contact, 2.02% conversion, and 88% of checkouts right today. Lost orders are the largest loss, and the one no report shows. The free check replaces each input with your own, with your permission.
Once Selva serves your shoppers, 20 of the 21 go live. Same definitions, counted from what shoppers actually did. Accessibility stays a release check.Once Selva serves your shoppers, the report runs every week. Same definitions, counted from that week’s sessions, each one linked.
Kept dollars per session, net of returns12 weeks
Shoppers Selva answeredHoldout: your store as it is
What it foundExamples, not results
61 front-pad returns last quarter were four-piston pads sold to six-piston cars.Changes the question Selva asks
Your wheel filter hid 6 black 17-inch wheels their maker lists for the JL.Changes what your selector shows
Your intake page says “50-state legal”. Its California approval covers the 3.5 EcoBoost, not the 5.0.Changes what you can ship to California
14 bridesmaid orders last month were promised a date the alteration window couldn’t meet.Changes when Selva says “order by”
Your Sleeves filter hid 9 dresses with three-quarter sleeves.Changes what your filters show
Your size chart and your product page give the sage chiffon two different waists.Changes which page you fix first
04The proof
What the store itself pays per 1,000 sessions. Solid is the floor, the outline runs to the ceiling. Lower is better.What wrong answers cost the store per 1,000 sessions, and where it comes from. Lower is better.
Fit returnsHidden fitsLost at a questionContactsUp to the ceiling
SelvaAsks only when the answer changes the verdict
Year, make and model badgeSays “fits” on any year, make and model match
Deep selectorMakes shoppers pick packages up front, from memory
Fit notes, one click awaySays “verify fitment” and links the notes
Floor: fit returns anchored at an estimated 4% of orders, with 8% of maker-listed fits hidden at every store but Selva’s. Ceiling: shopper errors at published values. Your own measured rates replace both. Selva is lowest in all 48 sensitivity runs.One typical apparel store at central inputs: a $125 order, 35% margin, 2.02% conversion, and 88% of checkouts right today. The sales Selva’s own questions lose are counted against it.
Key findingsSimulated
$44 per 1,000 sessions at the floor, against $142 for the year, make and model badge and $388 for fit notes one click away.
The floor sets the badge store’s misfits at the 4% anchor: 4.0 of every 100 orders. Selva’s design brings that to 1.4.
46 seconds against 28, counting successful sessions only. Count every wrong or failed cart, and Selva takes 15% less time per correct cart at the floor and 72% less at the ceiling.
Shoppers who leave at Selva’s question are counted against Selva. The deep selector, which asks up front, loses 0.29.
From Selva’s internal simulation of 300,000 auto-parts sessions, in its benchmark library of November 3, 2026. The floor anchors fit returns at an estimated 4% of orders; the ceiling keeps shopper errors at published values.
Key findingsSimulated
A party rests on 7.5 decisions, so small misses compound. For a party of five, 50 against 92.
The largest single cause of a miss is a question nobody asks: the wedding date.
$138.50 of every $165.34 per 1,000 sessions, after Selva pays for its own questions. Most of it is orders that were lost, not returns.
16 shoppers in every 1,000 sessions get one deciding question, and 0.24 of them leave. It is counted against Selva, beside the win.
From Selva’s internal simulation, built on its priors, calibrated so 87% of occasion and 88% of apparel checkouts are right today. The model prices Selva at 99 in 100 a decision, a cautious figure and not a measurement. Selva goes live on a store only when every mission passes on that store’s catalog, three runs of three. What remains in Selva’s row is shoppers who answer wrongly about themselves. The gate closes it with a way to tell that doesn’t rely on memory.
The same four designs on eight measures. Selva leads where the money is and pays for it in seconds and taps, printed here beside the wins. Open a column for its floor, its ceiling and what it is checked against.How many missions end right, of 100, at a typical occasion or apparel store and with Selva’s design. The dollar line is the apparel store’s. Open a column for the decisions it rests on and what breaks them today.
Shown in full: seven of the auto-parts set. The other fourteen are named below. Every category gets its own, in its own words, and every store sees all of its own before its replay.Shown in full: seven of the 21, worded for fashion as a worked example. The other fourteen are named below. Every category gets its own, in its own words, and every store sees all of its own before its replay.
A rotor badged “Fits” on every trim, when the maker lists it for one brake package.
The maker’s application guide
A skid plate the maker lists for the Gladiator that your vehicle filter never shows.
The maker’s application guide
A shopper types “jl front brake squeal” and gets wiper blades.
The part that fits the shopper’s vehicle
The right part in the cart with fit confirmed. Not just any cart.
A stopwatch, on correct carts only
An owner sure they have the base brakes, whose car has the performance package.
The vehicle’s true configuration
A 3.5-inch lift and stock-length shocks in one cart, with no warning.
Each maker’s compatibility rules
An AI assistant asked for pads for a 2019 GT picks the wrong brake package.
The same answer key, three runs out of three
Fit truth
Finding and promises
Time and effort
Machines, law and access
The free check answers two of these first, from your public pages: hidden fits, which decide the most, and false fits.The free check starts where fashion shoppers feel it first: the ready-by date and the whole party, read against your own chart and published times.
Get the free checkThe clock only ran for winners. The model’s time to a confident cart averages successful sessions only. The badge store’s fast time leaves out the 11% of shoppers it fails at the floor, and 39% at the ceiling. Selva’s time includes the hard cases it gets right.
Fast because it guesses. At a badge store, a shopper who doesn’t know the deciding fact guesses, and the badge says “fits” either way. Browsing with no vehicle set reaches a cart in 9.5 seconds, a cart that is often wrong.
It times the fit task, not the visit. The seconds are an expert’s pace through the fit task: taps, reading, and checking the vehicle. They leave out the rest of a real visit, like browsing and weighing options, which is why no real visit looks like 46 seconds.
Time per correct cart: every session’s time, with each wrong or failed cart counted at the 300-second cap, divided by the correct carts. The cap is the shortest a run is allowed, the most generous setting for the badge. Successful sessions enter at their simulated geometric-mean time. The real number comes from timed runs on real sites, not from the model.
Every decision has to hold. A mission ends right only if each decision on its path is right: the size, the date, the dress code, the color. At a typical occasion store’s 8.2% error a decision, one dress resting on three decisions ends right 77 times in 100.
A party multiplies them. A wedding party rests on the palette, the date and the shipping, plus one size for each member: 7.5 decisions for a typical party of 4.5. The same 8.2% now lands 53 parties in 100, and a party of five 50.
Selva goes live only when every decision passes its gate, three runs of three. Even at 99 in 100 on each, the model’s cautious figure, the dress lands 97 times in 100 and the party 93. A store can look fine decision by decision and still lose half its parties, and a lost party is up to five orders, felt as one event.
Simulated: (1 − e)D, with the typical occasion store’s e of 8.2%, set so 87% of its checkouts are right today, against Selva’s 1%. Your own sessions replace the model, with your permission.
05Going live
01Free
Selva reads your public pages at a human pace and measures the rate that decides most of the money, the fits your store hides, along with the fits it shouldn’t claim.
02Paid only if it clears its floor
Your own 90 days, replayed. Every return is matched to the verdict Selva would have given, and the prediction is written down before any shopper sees Selva.
03Against a holdout
Selva answers a share of your shoppers. The rest see your store exactly as it is today, so the difference is measured, not argued.
04Every week
The benchmarks become your report: 20 of the 21, with the same definitions, counted from what shoppers actually did.
01Free
With your written yes, Selva reads your product and policy pages against your own size chart, order-by windows and policy. Every finding is read twice.
02Paid only if it clears its floor
Your last 90 days of orders, replayed: order by order, the verdict Selva would have given.
03Against a holdout
Selva answers a slice of your traffic. The rest see your store exactly as it is today, so the difference is measured, not argued.
04Every week
The report runs every week: the same definitions, counted from that week’s sessions, each one linked.
Each step opens only when the one before it has proved itself.
Questions
06Where this goes
A store used to be a set of pages a shopper searched. Now the shopper asks a whole question, and more and more often an agent asks it for them. The part of the store that answers is where the sale is won or lost.
Selva starts where a wrong answer costs the most: things that have to fit. A part to a vehicle, a stove to a chimney, a dress to a body and a date. It is opening to its first stores now, and each one gets a test written for it before Selva answers a single shopper.
Requests go straight to Ken Navarro, founder.[email protected]267-265-9284 · Wayne, Pennsylvania