AI outreach optimization: how to improve reply rates with AI

Outreach optimization means changing one thing at a time and measuring the effect. In a year of our own LinkedIn outbound the levers that moved results were the list (intent signal lists replied at 54.6% against 21.7% for Sales Navigator searches), the opener (observations beat questions by 10 to 15 points) and follow ups (they lifted replies from 26.5% to 38.4%). Faster replies did not lift meetings. Sellinger's config optimization proposes changes daily, at most three per cycle, with auto approve off, so a person decides what goes live.

"Optimize with AI" is easy to promise and hard to do well, mostly because most of what gets changed does not matter. This guide ranks the levers by evidence, shows what an optimizer can and cannot do for you, and gives a test routine that stops you from reading noise as progress.

Which levers matter most?

We ran our own LinkedIn outbound for a year and recorded what changed. The details are in the Sellinger case study. Ranked by the size of the effect we saw:

A second client points the same way. In the ShortVibe case study, 16 of 19 booked calls came from intent lists, while title searches booked 1 in 225. These are two accounts, not a law, but they agree on the order: start with who you write to.

What does an optimizer actually do?

An optimizer reads how a campaign is doing and proposes changes. In Sellinger that is config optimization, and it works like this:

• It proposes changes to the campaign's custom directions, the first message, the follow up count and interval, flow sequences and more.
• It leaves some things alone. Product info and inclusion and exclusion rules are excluded from its changes, so it cannot rewrite what you sell or whom you target.
• It runs on a schedule. The default is daily, with at most three changes per cycle and a minimum confidence of 0.7.
• It does not act alone by default. Auto approve is off. Changes appear as proposals that you apply or revert.

So an optimizer is a fast reader of your data and a source of ideas, with a person on the switch. We publish no result for the optimizer, because we have not isolated one. Treat each proposal as a hypothesis, and use the routine below to test it.

How do you test a change without fooling yourself?

Most "improvements" are noise. Use this routine.

1. Change one thing. The opener, or the follow up gap, or the list. Not three at once.
2. Split the audience at random. Half gets the new version, half the old, in the same weeks, from comparable senders.
3. Decide the sample in advance. At least 100 people per variant for a reply rate, and more if the rate is low. With 50 per arm, three extra replies look like a ten point win.
4. Count the right outcome. For a first message, replies. For a campaign, meetings per 100 people messaged. And separate real replies from pleasantries: in the referral network case study, 1,751 of the first replies were "thanks for connecting", which inflates any reply rate.
5. Wait for the follow ups. A variant changes behaviour over weeks, so read results after the whole sequence has run.
6. Repeat once before you change your default. The calculator does the arithmetic.

A worked example of one test

Say you want to know whether an observation opener beats a question opener. You have 400 people on one list, and you split them at random into two groups of 200. Both get the same follow ups, from comparable senders, over the same three weeks.

• Group A, observation opener: 200 messaged, 86 replied (43%).
• Group B, question opener: 200 messaged, 68 replied (34%).

The gap is nine points, close to the 10 to 15 points we saw in our own data. With 200 per group, a nine point gap is unlikely to be chance alone, but it is not proof for your market either. So you repeat on a second list. If the gap holds, make the observation opener your default. Next, remove the weaker idea from the pool, and test the next lever. The numbers in this example are made up to show the arithmetic. They are not results.

Three things would make you distrust a test like this: groups that were not random (for example, the best leads all in A), a second change slipped in midway, or a reply count that includes pleasantries in one group and not the other.

What should you optimize, in order?

1. The list and the signal. Narrow the audience. Prefer people who just did something relevant over people who merely match a title. The post on segmentation and the one on intent signals explain how.
2. The opener. State one true observation about their work. See observation openers beat questions.
3. The follow ups. Add value in each, and stop cleanly. Follow ups add replies, and the share of negative first replies rises with each one: in our data it was 0.9% on the intro and 14.7% by the third follow up, which is the cost to weigh. The post on follow up copy shows what worked.
4. The offer and the ask. A smaller first ask, such as 15 minutes.
5. Timing and speed. Last, because it did not move meetings for us.

What should the AI not touch?

• Facts about your product. Keep them fixed, which is why Sellinger excludes product info from optimizer changes.
• Who you are allowed to contact. Inclusion and exclusion rules, suppression lists and opt outs are yours.
• Claims. No invented numbers, customers or quotes.
• Limits. Daily limits and quiet hours protect the account. An optimizer that raises volume to hit a reply count is optimizing the wrong thing.

A weekly rhythm

• Monday: read last week's numbers by list and by variant.
• Tuesday: read the optimizer's proposals and apply at most one.
• Wednesday to Friday: let it run, and read ten real conversations to see what people say.
• Next Monday: compare, keep or revert.

One change a week is slow on purpose. It is the only way to know which change did anything.

A checklist

• [ ] I know my reply rate and meetings per 100 messaged, by list.
• [ ] I change one thing at a time and test against a control.
• [ ] Each variant has at least 100 people before I read it.
• [ ] Auto approve is off, and I apply optimizer proposals one at a time.
• [ ] Product info and exclusion rules are not up for change.
• [ ] I read real conversations every week.