How to do cold outreach on LinkedIn (without getting restricted)

LinkedIn's own restriction page names three triggers, and only one of them is about how much you send. The other two are about how recipients reacted — invitations ignored, left pending, or marked as spam — and whether the pattern looks like an automation tool. That reframes the whole problem. Staying unrestricted is mostly a targeting discipline with a pacing discipline bolted on, and a procedure that gets the order wrong will keep tripping the same wire while carefully obeying a daily counter.

Restriction is a verdict on your targeting, not just your volume

The three triggers, quoted from LinkedIn's page on types of restrictions for sending invitations, are that you have sent many invitations within a short amount of time; that many of your invitations have been ignored, left pending or marked as spam by the recipients; and that if you send an excessive number of invitations and LinkedIn suspects the use of an automation tool, it may suspend or restrict the account.

Notice what that means for diagnosis. If you are restricted at a volume that was fine last month, the change may be entirely on the recipients' side. Your send rate is a number you control; your ignore rate is a number your list controls. A team that responds to a restriction by lowering the daily cap while keeping the same list is treating a symptom on the wrong side of the equation.

LinkedIn's advice on the same page is short and unfashionable: only send invitations to people you know and trust, and provide context by adding a personalised message explaining how you know the recipient or why you would like to connect. That is a stricter standard than any cold-outreach programme actually meets, and being honest about that gap is the beginning of a sane procedure rather than a reason to ignore it.

The account has to look like a person before it acts like a salesperson

An account that has produced no activity for a year and then sends its first fifty invitations in an afternoon has a shape. Fix the shape first, because none of the later steps compensate for it.

Complete the profile. Photo, banner, a headline that states what you do rather than a job title, a populated experience section, and enough skills that the profile does not read as a shell. LinkedIn lists an updated profile with a photo among its own suggestions for avoiding restriction.

Produce non-outbound activity for a week or two. Comment on posts in your space. Publish something short. Accept inbound invitations. The point is not reach — it is that the account has a history that is not exclusively one-directional.

Decide what the profile has to prove. Someone deciding whether to accept your invitation will look at your headline and your most recent activity, in that order, and usually nothing else. If those two surfaces do not make the reason for the connection obvious, the note is doing work the profile should have done.

Fix the About section last, not first. Almost nobody reads it before accepting. It matters after acceptance, when the conversation is deciding whether you are credible.

Day 1 is not day 21: the ramp, and what a real one looks like

The single most reliable way to get flagged is to open at the ceiling. A ramp is not caution for its own sake — it produces the gradually increasing activity curve that a returning human user produces, and it gives you a week of acceptance data before you are spending at full rate.

A workable shape, expressed as a procedure rather than as a promise:

1. Days 1 to 3: around five invitations a day, all to people where acceptance is close to certain — former colleagues, existing customers, people who have engaged with your posts.
2. Days 4 to 10: grow gradually, and hold the increase if acceptance rate falls. This is the first week where the number is telling you something.
3. Days 11 to 21: approach your intended steady state, still spread across the working day rather than clustered.
4. After day 21: hold. Do not treat the ceiling as a target to beat.

Autocloz implements exactly this shape rather than leaving it to a rep's discipline on a slow Friday. The warmup ceiling starts at a floor of 5 actions on the account's first day and grows on a smoothstep curve to the configured daily target over 21 days by default, so growth accelerates through the middle of the ramp rather than climbing linearly. On Saturday and Sunday in the account's own timezone the allowance drops to 30% of the weekday figure, because an account shipping a full day's invitations at noon on a Saturday is a giveaway. And every successful action pushes the next one out by the configured minimum gap plus a random 60 to 300 seconds, so the interval histogram is irregular rather than periodic — a rigid gap of exactly one minute is as detectable as no gap at all.

When a gate blocks a step, the deferral carries a named reason rather than failing silently: linkedin_warmup_or_weekend_cap with today's count and today's ceiling, linkedin_weekly_invite_cap with the used total and the cap, linkedin_pending_invite_cap, or linkedin_burst_pattern_detected when five actions have landed inside a ten-minute window. A defer you cannot read is a defer you cannot act on.

Who to invite: the one question that predicts acceptance

Before an invitation goes out, answer this: would this person be able to say, in one sentence, why I appeared?

Not "would they benefit from the product". Not "do they match the ICP". Whether the connection has a legible reason from their side. That question predicts acceptance better than seniority, industry or company size, and it is the only lever that moves LinkedIn's second trigger.

Legible reasons, roughly in descending order of strength:

• You have interacted. They posted, you commented substantively, they saw the name.
• A mutual connection who would vouch. Second degree with a real overlap, not a shared connection you both happen to have collected.
• A shared context. Same former employer, same conference, same customer, same community.
• An adjacent role in an obviously related market, where connecting is normal professional behaviour.
• Nothing but a title match. This is the category that generates ignored invitations, and it is also the category most list-building tools produce by default.

If more than about half your list falls into the last category, the pacing conversation is premature. Narrow the list. Building that list properly is its own discipline, set out in how to build a targeted lead list.

Pending invitations are a budget, and yours has a ceiling

Outstanding invitations accumulate. Every one that is never acted on sits there contributing to the trigger LinkedIn ties to its longest restriction — the one where its guidance is to wait up to a month.

Two things follow. First, you need a withdrawal policy rather than a withdrawal event. Autocloz can withdraw invitations still pending after 14 days, but be clear about how: it is an operator-triggered sweep, not a scheduled job, and nothing runs it for you. The provider-side rescind is documented in the code as forthcoming, so today the sweep updates the ledger. Batch it deliberately rather than rescinding fifty invitations in a tight loop is itself a burst pattern. Second, the withdrawal has a cost LinkedIn states directly: after you withdraw an invitation you cannot resend one to the same recipient for up to three weeks. A single bulk cleanup therefore locks a slice of your list out of reach for most of a month, which is a real reason to do it continuously instead.

There is also a hard stop worth having. Autocloz refuses to send further invitations once 1,000 are pending on a sender, deferring for twelve hours with a linkedin_pending_invite_cap reason, so the withdrawal sweep has room to drain when you run it rather than the account discovering the limit from LinkedIn's side.

Both the pending count and the withdrawal schedule are visible per sender on the LinkedIn channel page, which is the surface to check before raising anyone's daily figure.

The governor that ties all of this together is acceptance rate. When the trailing accept rate on a sender falls below 30%, Autocloz halves the effective daily ceiling for that dispatch — and it deliberately does nothing when fewer than ten invitations have been sent in the window, because throttling on a statistically meaningless sample punishes a healthy sender for a slow week.

What happens after acceptance, and the pause that has to happen first

An accepted invitation is not a conversation. It is permission to have one, and the most common way teams waste it is by messaging within the minute.

Wait. Two or three days is a reasonable default, and the reason is not etiquette — it is that an instant message after acceptance is the clearest possible signal that no human was involved on your side. LinkedIn's third trigger is about suspected automation, and a zero-latency follow-up is the cheapest evidence you can hand it.

When you do write, the first message after acceptance should not pitch. It should do one of three things: ask a question you genuinely cannot answer from their profile, share something specific to their situation, or state plainly why you connected. Anything that reads as a template with a merge field is worse than nothing, because it converts a neutral connection into a negative impression on an account you now cannot un-connect from cheaply.

Autocloz enforces the ordering structurally rather than trusting sequence design. A linkedin_dm step checks for a prior invite touch on that lead recorded as accepted or replied, and skips with "DM steps require a prior accepted linkedin_invite for this lead" when there is none — so a sequence cannot accidentally message people who never connected. Separately, if the lead has replied on any channel in the previous fourteen days, the LinkedIn step stops with lead_replied rather than firing, because a templated follow-up landing after a human reply is the fastest route to being reported.

A seven-step LinkedIn sequence with every gate written down

A concrete cadence, with the gate that each step passes through. Days are from enrolment, and every step is subject to the pacing layer above.

1. Day 0 — profile view. Costs nothing, generates a notification, and gives the invitation a small amount of prior context.
2. Day 1 — invitation, with a note under 200 characters. Subject to the warmup ceiling, the weekly cap, the burst check, the pending-invite cap and both deduplication gates.
3. Day 4 — nothing. This step is deliberately empty. Pending is not a failure state at four days.
4. Day 8 — if accepted, the first message. Not a pitch. Gated on the accepted-invite check.
5. Day 12 — if accepted and no reply, a second message from a different angle. Different angle means a different reason to respond, not the same reason rephrased.
6. Day 15 — if still pending, stop the LinkedIn branch. Do not escalate on the same channel.
7. Day 15 onward — the invitation is eligible for withdrawal, but only when you run the sweep. Nothing frees the pending budget on its own.

Two gates sit underneath all seven and are easy to forget. Autocloz refuses to touch a lead that another LinkedIn sender in the same workspace engaged within the last 90 days, recording cross_sender_dedup, and refuses when another lead at the same company was contacted within the last 30 days, recording cross_company_dedup with the offending touch id. Those exist because the compounding version of a spam report is five reps converging on one account in one month, and no per-rep daily cap prevents it.

Running the LinkedIn branch alongside email and calls in one sequence, rather than in a separate tool, is what makes step 6 possible at all — you need somewhere for the prospect to go when the LinkedIn branch stops. Autocloz's free plan covers 5 users and 10 mailboxes across all five channels with a LinkedIn allowance of 50 actions a day, so a paced branch fits inside it — start free. The multi-channel sequence builder is where the branch and its stop condition are configured.

Diagnosing which of the three triggers you actually hit

Restriction arrives without a diagnosis, so build one. Each trigger has a different fingerprint and a different remedy, and applying the wrong remedy is how teams get restricted twice.

Fingerprint of the volume trigger: the restriction follows a step change in activity — a new tool, a new rep, a ramp that skipped its middle. Acceptance rate is unchanged. The remedy is pacing, and it works.

Fingerprint of the reaction trigger: acceptance rate has been falling for a fortnight, pending invitations have been climbing, and the restriction arrives at a volume that was previously fine. The remedy is the list, and lowering the cap only delays the next restriction.

Fingerprint of the automation trigger: activity is metronomic, or it originates from a session that does not match the account's normal device and location, or actions occur at hours the human never works. The remedy is jitter, session consistency, and honestly reconsidering the tool.

Instrument three numbers weekly so the fingerprint is readable before the restriction arrives: acceptance rate by segment, outstanding pending count, and the distribution of gaps between consecutive actions. The third one is the one nobody tracks and it is the one that distinguishes trigger three from trigger one. The wider version of this — what each channel's adjudicator can observe and how its refusal looks — is worked through in multi-channel outbound without getting banned.

What this procedure cannot prevent, and what Autocloz does not do

The honest limits, because a safety procedure that claims completeness is not a safety procedure.

Nothing here makes enforcement impossible. LinkedIn's User Agreement prohibits using bots or other unauthorised automated methods to access the Services, add or download contacts, or send or redirect messages, at clause 8.2.13, and prohibits developing or using software to scrape the Services at 8.2.2. Any tool that acts on your behalf runs against that language regardless of how carefully it paces, and nobody selling LinkedIn automation can honestly tell you the risk is zero. Careful pacing lowers probability. It does not confer immunity, and a permanently restricted account takes its connection graph with it.

Every pacing figure in this post that comes from Autocloz — the 80-invitation week, the 5-action first day, the 21-day ramp, the 30% weekend factor, the 1,000 pending ceiling, the 14-day withdrawal, the 30% accept-rate floor — is a product default, not a LinkedIn rule. LinkedIn publishes no invitation number at all. Treat any tool that quotes you a platform limit as a number to the decimal place as quoting its own configuration back to you.

And Autocloz specifically. It cannot see whether a message was read, because LinkedIn exposes no read receipt to an integration. It cannot detect a soft throttle, where invitations still send but stop surfacing — that shows up only as acceptance rate falling, which is why the accept-rate governor exists and why it is a lagging signal rather than an alarm. It does not sell LinkedIn accounts, Premium seats or InMail credits; the subscription is yours and bought from LinkedIn. When the LinkedIn session is unavailable the actions degrade to a loud unavailable result rather than a fake success, which is correct but means a disconnected cookie stops the branch rather than queueing it indefinitely. The credit-versus-standing trade-off between an invitation and a paid message is a separate decision, worked through in InMail against the connection request, and the underlying session integration is documented on the LinkedIn integration page.