Three layers of numbers
Delivery says the message arrived. Engagement says someone did something: clicked, replied, entered a code. Outcome says it was worth sending: an order, a booking, a renewed subscription. The delivery report gives you the first; the other two come from your own systems, and they are the ones to optimise.
| Layer | Question | Where the numbers live |
|---|---|---|
| Delivery | Did it arrive? | The delivery report: statuses, failure reasons, price per message |
| Engagement | Did anyone act? | Your web analytics, your reply inbox, your code verification |
| Outcome | Was it worth it? | Your orders, bookings and renewals |
Before the send, write down which outcome counts and how long after the send it may happen: the attribution window. Seven days for a sale or 24 hours for a flash offer are choices, not rules. What matters is keeping the same window from one campaign to the next, so the figures stay comparable.
Reading the delivery layer
Delivery rate is delivered messages divided by messages handed to carriers. Leave rejected messages out: they were refused before reaching a carrier, are never billed, and are a problem to fix in the file. Then read the rate per destination, because receipts differ. Most destinations return handset receipts, where delivered means the phone got the message; a few give carrier acknowledgement only, where sent is the final positive state. Averaging the two kinds drags the rate down without anything being wrong; each country page says which kind applies.
Failures are grouped into about twelve plain reasons, each with the raw carrier code, explained in the guide to why SMS fail. Two shares are worth tracking from send to send: unknown numbers, which measure how fast the list ages, and carrier filter blocks, which measure how operators receive your content and sender.
To join delivery with your own data, export the report as CSV or JSON: message id, number, status, carrier code, timestamps and price for every message. Per-message records are kept for 90 days and campaign summaries indefinitely, so export what you want to keep. API senders can skip the join: a reference of up to 128 characters, such as an order or user id, is echoed in responses and in webhooks, and the message.delivered and message.failed events update your own records as they happen.
Links you can count
Use a link on your own domain with a campaign parameter, for example ?c=sept-sale, and count visits and orders on your side. Avoid public shorteners: they are the most filtered pattern in SMS and they hide your brand. A short path on your own domain costs the same characters.
- Short in the message, complete on the server. Send
acme.example/s/bootsand let your server redirect it to the full address with every analytics parameter you use. The message spends a few characters; your analytics gets everything. - One link per recipient, when you need it. A column of unique codes in the import file turns
acme.example/r/{code}into a personal link, so a visit identifies the recipient rather than only the campaign. Keep the codes the same length, so the longest rendered row stays predictable, and never put a name, email address or phone number in a URL. - Count people, not hits. Link previews and automated scanners can fetch a link without anyone tapping it. Unique visitors are sturdier than raw hits, and outcomes are sturdier still.
- Test the link on a phone before the send. A broken link wastes the whole campaign, and a sent message cannot be recalled.
Cost per outcome
Campaign cost divided by outcomes. At an illustrative rate of USD 0.01 a message, a campaign of 50,000 costs USD 500; 400 orders is USD 1.25 per order. That figure, not the delivery rate, is what you compare with email or ads, and it is usually the reason SMS stays in the mix.
Use the billed cost, the sum of the price column in the export, rather than list size times rate. It already accounts for segments, for each destination's rate, for failed messages, which reached a carrier and are billed, and for rejected ones, which are not. Segments are where estimates go wrong. The same campaign in two segments costs USD 1,000, or USD 2.50 per order; a curly apostrophe that pushes the text into three Unicode segments makes it USD 3.75. The GSM-7 vs Unicode guide shows how to stay in one.
Three ratios tell the story together: cost per delivered message, cost per unique click and cost per outcome. If you know the margin on an order, the last one answers the question that matters most: did the campaign pay for itself?
What SMS actually caused
Attributed outcomes overstate a campaign, because some of those customers would have ordered anyway. A holdout group measures the difference. Before the send, set aside a random share of the list, say 10%, and do not message it. Afterwards, count every order placed by each group within the attribution window, whether or not anyone clicked.
An example with made-up figures: a list of 50,000 is split into 45,000 recipients and 5,000 held out. The recipients place 540 orders, 1.2%; the holdout places 30, 0.6%. The campaign caused the difference, 0.6% of 45,000, which is 270 orders. At USD 0.01 a message the send cost USD 450: about USD 0.83 per attributed order, but about USD 1.67 per order it actually caused. The second figure is the honest one to compare with other channels.
A holdout gives up the outcomes the held-out group would have produced, so keep it small, but large enough that its count is not mostly chance; the check further down applies to it too.
Opt-outs are a metric
Opt-out rate is opt-outs divided by delivered messages, per send. Count both ways out. Replies such as STOP need a number that can receive them: alphanumeric senders are one-way, while a rented number with inbound passes replies to your webhook, as described under two-way SMS. Opt-out links on your own domain are counted on your side, like any other link.
A jump after a change of cadence or tone is the audience telling you something. Compare each send with the same list's previous sends rather than with a figure from elsewhere: if a list that usually loses 30 people per send loses 120 after you moved from monthly to weekly, the cadence is the answer. Opt-outs are permanent for that list, so the cost of a careless campaign is paid on every later one. Complaints you never see work the same way: they reach operators, and a rising share of filter blocks in the delivery report can be their echo.
Comparing two sends: A/B tests
Change one thing at a time: the hour, the first five words, the offer. Split the list randomly, not by country, because sender rules and delivery differ per destination and would drown the effect. Tag each half with a different campaign so the reports line up. In practice, an A/B test that can be trusted runs like this:
- Pick the variable and the deciding metric before sending. The metric should be an outcome, or at least a click; a delivery rate cannot tell two texts apart.
- Split at random. In a spreadsheet, add a column of random numbers, sort by it and cut the list into equal parts. Alphabetical order, sign-up date or country all smuggle in a second difference.
- Keep everything else equal: same sender, same window, same day, unless timing is the variable.
- Tag each group with its own
campaign, up to 64 characters in the API, or send each group as its own campaign in the panel. - Use the same attribution window for both groups.
- Test on a sample, then roll out. Two test groups of 10% each, then the winner to the remaining 80%, keeps most of the list on the better version.
Is the difference real?
Small differences between two groups are often chance. A rough check for groups of equal size: add the two outcome counts, take the square root and double it. If the gap between the counts is larger than that, it is probably real; if not, call it a draw and test something bolder. With 200 orders against 240, the gap is 40 and the threshold about 42: a draw. With 200 against 260, a gap of 60 clears a threshold of about 43. The check suits counts of fairly rare events such as orders or clicks; for decisions that carry real money, use a proper significance test.
What to report
A report someone can act on fits in one table per send, set next to the previous send to the same list:
| Figure | Source |
|---|---|
| Rows imported, invalid, duplicates merged | Import report |
| Sent, delivered, failed by reason, expired, rejected | Delivery report, per destination |
| Encoding and segments per message | Composer or API response |
| Cost | Sum of the price column in the export |
| Unique clicks and replies | Your analytics and your inbox |
| Opt-outs and opt-out rate | Replies and your opt-out page |
| Outcomes, attributed and caused | Your systems and the holdout |
| Cost per outcome | Arithmetic |
Add the context that explains the numbers: the text, the sender, the window and the one thing you changed. Six months later, a figure without its cause is trivia. For the delivery rows, see reading a delivery report in two minutes; the rest is yours to keep.
Something unclear or out of date? Tell us on the support page; corrections are folded back into the guide and the update date changes.
