A bag seller I watched last winter had a bot that answered everything correctly. Price, colours in stock, delivery time, return window. All accurate, all inside ten seconds. Her sales did not move.
I read a week of those conversations. They were polite, they were fast, and nearly every one ended the same way: the bot wrote "let us know if you'd like to order", the customer wrote "thanks", and that was that.
Nobody was ignored. Everybody was answered. The money still walked out the door.
That gap is the thing worth talking about, because answering faster and selling more are two different projects, and most sellers only ever finish the first one.
Conversion isn't one number. It's a chain of four.
When a seller tells me their DM conversion sits at 4%, that number is hiding four separate steps, and usually only one of them is broken:
- Reached — did the person get a real answer while they still cared?
- Qualified — did you learn what they actually want, plus size, city, budget?
- Matched — did you put one specific product in front of them instead of a catalogue?
- Paid — did you ask for the order and hand them a way to pay right there?
An AI chatbot that only fixes step one lifts your reply rate and nothing else. That is exactly what happened to the bag seller: perfect at step one, invisible at steps two through four. If you want to know what the whole chain should add up to, we published the real benchmark for DM conversion rates from live seller data.
Most of your inbox is not shopping
Here is the uncomfortable part. When we looked at first-contact messages across sellers, roughly 15% were real buyers. The rest were price-checkers, wholesale hunters, people who wanted a job, people who wanted a collab, and a healthy number who just said "hi" and vanished. The full breakdown is in our anatomy of first-contact DMs.
A human seller feels this and gets tired. By message forty their answers get shorter and colder, and the good buyer who arrived at message forty-one pays for the mood of the previous forty.
An AI assistant does not get tired, which sounds like the whole benefit. It isn't. The real benefit is that it can sort. The wholesale question deserves a different answer than "does this fit a wide size 40 foot". One gets a polite policy line, the other gets a sale.
Sorting is where conversion actually starts. If your bot treats all 100 messages identically, you have automated the tiredness, not removed it.
One good question beats five polite ones
Bad assistants interrogate. Name, then city, then budget, then colour, like a form that developed a personality. Buyers bail somewhere around question three.
Good ones ask the one question that unblocks the recommendation, and only that one.
For clothing it's usually height and usual size. For shoes, foot length in centimetres, because everyone is optimistic about their size. For skincare, skin type. For electronics, what they'll use it for and their ceiling on price.
Notice what these have in common: the answer changes which product you show. If an answer wouldn't change your recommendation, don't ask for it. Collect the rest at checkout, where people expect to type things.
Ask it inside a suggestion, not before one
"What size are you?" is a chore. "The 42 is what people usually take for a wide foot, and we have it in black and tan. What's your foot length in cm so I can confirm?" is a conversation. Same question, but the buyer already has something to picture.
Speed is a conversion lever, not a service metric
Sellers file reply speed under customer service. It belongs in sales.
The window where a DM buyer stays warm is short. Most of them are messaging two or three sellers at once, at night, from bed, with the app open. Answer inside a couple of minutes and you are talking to someone who is still comparing. Answer at 11 the next morning and you are talking to someone who has already paid a competitor and is now just being polite to you. We went through the numbers behind that in the five-minute window between response time and sales.
This is the one place where automation wins with no cleverness required. Nine seconds against sixty minutes is not a small edge. It is the difference between joining the conversation and reading the transcript afterwards.
Suggest a product. Don't recite the catalogue.
The fastest way to kill a warm DM is to send six options.
Choice paralysis is real, and it is worse on a phone, in a chat window, with photos still loading. Two options with a clear reason to pick one convert better than six with no reason at all.
What a good AI assistant should be doing at this step:
- Name one primary pick and say why it fits what they just told you.
- Offer one alternative, usually cheaper or a different colour, so it feels like a choice rather than a push.
- Say the price out loud. Hiding price to "get them into a conversation" is a tactic from 2016 and buyers hate it.
- Mention stock honestly. "Last two in that size" only works while it's true, and buyers can smell when it isn't.
- Attach a photo of the actual item. Not a mood board.
The assistant can only do this if it knows your catalogue properly. Feed it the sizes, the stock, the fabric, the shipping cost to different cities. A chatbot that knows your FAQ but not your inventory can chat. It can't sell.
The payment link is the conversion step
If you take one thing from this: the sale happens where the payment link appears.
Count how many of your conversations ended with something like "let us know when you decide". Every one of those handed the decision back to a person who was, thirty seconds earlier, ready. You had them at yes and you asked them to go think about it.
Replace that ending with a checkout. A payment link, a card-transfer instruction with an order confirmation, a cart the assistant fills for them. Anything that turns "I'll think about it" into a tap.
The mechanics matter more than people expect:
- Send the link in the same message as the confirmation, not after another round of questions.
- Restate exactly what they're buying. Item, size, colour, total with shipping. Surprise costs at checkout are the most common reason a filled cart dies.
- Keep the link alive for a day or two, and follow up once if it goes unclicked. Once. Two follow-ups reads as desperate, three gets you blocked.
- Someone who clicked and dropped off is a different customer than someone who never clicked, and deserves a different message.
Sellers who change nothing else but this step routinely see their biggest single jump, because it plugs a leak at the bottom of the funnel where every visitor has already been paid for.
Know when to get out of the way
Some conversations should not be automated, and pretending otherwise costs more than the sales you save.
A customer whose order arrived broken. A wholesale enquiry worth thirty times your average basket. Anyone who asks the same thing twice, which almost always means your assistant answered a different question the first time.
A good setup spots those and pulls a human in, with the whole thread already read, so the seller doesn't make the customer repeat themselves. The customer usually can't tell where the switch happened. That's the point.
What to actually measure
Reply rate becomes a vanity metric the moment you add automation. It goes to 100% on day one and then tells you nothing.
Watch these instead, weekly:
- Qualified conversations as a share of total DMs. This tells you whether your sorting works.
- Payment links sent per qualified conversation. Most sellers have never looked at this number, and it is usually shockingly low.
- Links clicked, then paid. The gap between those two is a pricing or shipping problem, not a chat problem.
- Handoffs per hundred conversations. Rising means your assistant's product knowledge has gone stale.
Four numbers. If conversion drops, one of them moved, and you know which thing to fix instead of rewriting your whole assistant on a hunch.
If you'd like this running on your own inbox, that's what Vardast does: it connects to your Instagram, WhatsApp, Telegram and website, learns your products and prices, qualifies buyers in your own tone, suggests the right item, and sends the payment link before the moment passes. Then it hands you the conversations that genuinely need a human.
Your bot doesn't need to get smarter. It needs to stop stopping one message before the sale.