How to diagnose and fix declining LinkedIn post reach

Declining LinkedIn post reach is a series, not a mood. Take the posts you can still open, keep one format, and compare the middle of a recent window with the middle of an earlier window. If that middle impression count is down, you have a decline worth diagnosing. One quiet Tuesday, or the week after a post that strangers re-shared, is not that series.

An impression is a display. LinkedIn's combined post analytics defines it as the number of times a post was shown on screen, and it says the number is an estimate. Members reached is the count of distinct members and Pages. The same screen splits displays into in network (people who follow you or are connected to you) and out of network (everyone else). How retrieval and ranking assign those displays is the LinkedIn algorithm guide. The actions LinkedIn has tied to fewer displays — bait, pods, automated comments, generic text, website bots — are catalogued in algorithm penalties. This page is the measurement in front of that catalog: which number moved, which comparison is fake, and which single change the next three posts are allowed to test.

Boykli drafts from a note, a link, a file, or a Voice Agent transcript, and it publishes an approved draft to the personal profile you connect. It does not import the discovery block, label a decline, boost a post, join a pod, leave a comment, or message anyone. The Algorithm page is a reading of LinkedIn's public notes. The impression number still lives on LinkedIn.

Logged-in visitors can browse Radar and Algorithm with no plan. Writing, carousels, Voice Agent, publishing, and scheduling need a paid seat. BYOK is $29 a month. Managed is $88 a month. Prices are the public list on boykli.com as of 27 October 2026.

Four numbers people call reach

Write these down as separate columns. A fix for one column does nothing for the others.

Number	Where it comes from	A decline in this column means	What it does not mean
Impressions	Displays on screen. LinkedIn calls the count an estimate	Fewer displays than the comparable posts in the earlier window	That the account is restricted, or that strangers stopped caring. You have not checked those yet
Members reached	Distinct members and Pages who got a display	Fewer people saw the post, even if some of them saw it more than once	That each viewer spent less time. Dwell is not this column
In-network vs out-of-network share	The discovery split on the same analytics screen	The mix of followers and strangers changed. A higher in-network share with fewer total displays is distribution staying closer to people who already follow you	A published penalty percent. LinkedIn has not printed one
Reactions, comments, and reposts, divided by impressions	A ratio you compute. LinkedIn shows the counts. It does not publish this ratio as a reach grade	The people who did see it acted less often. That can happen while impressions rise	That retrieval handed the post to fewer members. Reach and response are different incidents

What a good version of that ratio is in 2026, and which public average it is allowed to sit next to, is LinkedIn analytics benchmarks. One person can produce several impressions. A rising impression count with a flat members-reached count is repeat displays, not a new audience. The reverse belongs in a decline audit. Impressions down harder than members reached means each viewer produced fewer displays. Members reached down harder than impressions means fewer people saw the post more often. Those are different repairs.

Sponsored impressions are a different meter. Campaign Manager counts ads. Do not add them to the organic discovery block and call the sum a content decline.

False declines that waste a month

Most "my reach died" spreadsheets fail the comparison before they fail the writing.

What the sheet shows	Why it is not yet a decline	What to do instead
The last three posts versus your best post of the quarter	The best post is an outlier. Regression to the middle looks like a crash	Median of at least eight posts of the same format in each window. If you do not have eight, the sample is too thin to call
Text this month versus documents last month	Formats are different objects. Long dwell is a percentile inside similar posts, not one second-count across an image and a video. Zhang, Kothari, and Tiwana use 30 seconds as the illustration of that failure	Hold format still. A format test is a different experiment, covered in video vs carousel vs text
Today's post versus last month	Combined analytics is described as complete through yesterday. A morning post is an incomplete row	Drop anything newer than yesterday
Impressions per follower fell, while absolute impressions rose	The denominator grew. Followers who never open the app cannot produce a display. Members reached is the closer headcount	Keep absolute impressions and members reached. Do not grade the month as impressions divided by followers
The reaction ratio rose while impressions fell	A smaller set of viewers reacted more. That can be a sharper post shown to fewer people, or a pod that stopped and left only the people who meant it	Read both columns. Do not "fix reach" by adding a closer that asks for a reaction
The monthly total fell because you posted four times instead of twelve	Volume changed. The per-post middle may be flat	Compare medians per post, and write the post count next to the window so you do not call a quieter calendar a penalty
The week you left a pod	The early likes were the agreement. When the subsidy stops, the count drops to what strangers did	That drop is the measurement becoming honest. The month after is growing without pods
A company Page set beside a personal profile	Different audiences, different analytics. Boykli queues only to the personal profile you connect	Diagnose the surface you actually publish on

A clock change can also fake the story. Moving the slot and rewriting the hook in the same week makes the impression delta unreadable. The hour itself is chosen in best time to post on LinkedIn. The weekday pattern is best day to post on LinkedIn. Neither page is a recovery ritual. They are variables you hold still while you test the text.

Early activity is a third variable people bolt on. Popularity percentiles refresh within minutes, so the first replies can move which later sessions retrieve the post. Minute 60 is not a published switch, and a decline across eight posts is not explained by missing one sitting. The boundary is the first-hour note.

The diagnosis, in one sitting

Run this on a personal profile. Budget about 90 minutes. Do it once. Then change one thing.

Confirm the account can still publish (about 10 minutes). If LinkedIn shows a restriction, this is not a content decline. Automated activity tells you to disable the software or extension that acted on the website. The feature returns at the time on the notice. Until then, new posts do not display because they do not publish. A tool that asked for the LinkedIn password or the session cookie is in that class. An application you approved on LinkedIn's OAuth screen, with the scopes openid, profile, email, and w_member_social, is a different class: it can send a post you already approved. It does not browse the feed, and it does not raise the impression count. That split is automated LinkedIn posting.
Mark posts that disappeared (about 10 minutes). A removed post is a policy outcome, not a low median. Professional Community Policies cover spam, untargeted or repetitive promotion, and artificial engagement. Leave removed posts out of both windows. A new hook does not put a removed post back.
Build two windows of the same format (about 25 minutes). Pick text, single image, document, or native video, and stay inside it. Window A is the earlier stretch. Window B is the recent stretch. Each needs at least eight posts that are still on the profile and older than yesterday. Copy date, impressions, members reached, in-network percent, and out-of-network percent. Sort impressions. The median is the middle value, or the average of the two middle values when the count is even. The best post in each window can stay on the sheet. It does not set the baseline.
Write the post count beside each median. Twelve posts at a median of 800 impressions is a different month from four posts at a median of 800. The reach question is the median. The volume question is the count. Do not add the impressions and call the total a distribution score.
Tag each Window B post with one action, or with "specific." Specific means a stranger can repeat a rule, a number you measured, a sequence, or a constraint without reacting. Use one tag from the penalties page: engagement-bait closer, pod, automated comments, generic or recycled text, mismatched video, or repetitive promotion. If the closer is "comment YES," tag bait even when the paragraph above it is sharp. "The algorithm turned on me in March" is not a tag.
Read the split, not the vibe. Four patterns are allowed:
- Median impressions down, in-network share up. Distribution stayed closer to people who already follow you. On 4 June 2026, LinkedIn said content that looks generated and lacks a clear perspective is less likely to travel beyond the immediate network. On 15 September 2026, views of content classified as AI slop were down 40%. That 40% is LinkedIn's figure for the class its classifier labeled. It is not a tax on your account, and it is not a score you can look up.
- Median impressions down, in-network share about the same, and you also changed the hour, the format, or the weekly count. You do not have a clean cause. Freeze those variables before you rewrite the voice.
- Median impressions flat, members reached down. Fewer distinct people, more repeat displays among them. That is a narrower audience, not proof of a hidden flag.
- Median impressions flat or up, and only the reaction ratio fell. That is a response problem. A post that already contains the point, against a closer that withholds the point until someone comments, is the comparison in engagement versus value.
Stop if the sample is thin or the windows mix formats. Eight and eight, same format, both complete through yesterday. Anything smaller is a note, not a diagnosis. Do not average a document with a text post to manufacture a trend.
Pick one change for the next three posts. The order follows severity: website automation, then the pod or automated comments, then bait closers, then generic text. Keep format, clock, and length band where they were in Window B. The length band is a writing choice under a 3,000-character ceiling, not a recovery lever. The public character studies are in LinkedIn post length. A link in the body versus a line that says the link is in the comments is also not the first lever. LinkedIn has not published a link tax. Where 2026 samples put those placements is links in the post versus the first comment.

Hold the next three posts to one change

Create posts and Voice Agent sit on a paid seat. BYOK is $29/month with your keys. Managed is $88/month with platform models and monthly caps. The queue does not read your impressions or add a recovery boost.

boykli.com →

What you are allowed to change

Pattern you measured	The one change	Hold still	Leave this alone
Restriction notice	Disable the website bot. Wait for the time on the notice	No comeback post through a second extension	A pod bought so the profile looks alive while the feature is off
Posts removed	Stop the named policy act. Publish a new object only when you can sign it	Leave the removed URL out of the median	The same promo, rewritten, treated as a reset
In-network share up, median displays down, and the text could sit on any profile	Replace the next three with a receipt: the incident, what you refused, the mechanism, and what a stranger can repeat. That shape is the thought leadership framework	Format, clock, and weekly count	A topic-only prompt ("write five viral posts about leadership"). The June 2026 note says using a model to help write is acceptable when the perspective is still yours. The topic-only prompt is the other case
Pod or bought early likes in the earlier window	Leave the agreement off. Judge the new window only after the subsidy is gone	The topics you can actually sign	A new pod "to get the reach back"
A bait closer on an otherwise specific post	Put the useful object in the post. Ask for a fact only after the answer is already there	The incident you were already writing	"Comment the keyword and I will send the doc"
Medians flat once the outlier is removed	Do not treat the sheet as a decline	The cadence that produced the flat median	A format swap, a new hour, and a new hook in the same week
You changed the hour, the format, and the hook together	You cannot see a cause. Return to one clock and one format for three posts	The note you were going to publish anyway	A hashtag argument. Retrieval matches embeddings, including topics that do not share words. A hashtag block is not a published impression switch

Tim Jurka's 12 March 2026 note named "Comment 'Yes' if you agree" as engagement bait the feed would show less of, set against an actionable breakdown or a specific perspective. There is no public percent on that sentence. Relevance help says algorithms filter or taper low-quality and unsafe content. Taper is a distribution description. It is not a shadowban flag in the analytics UI.

Deleting old posts to reset distribution is not in these notes. A restriction has a time. A removed post stays removed. A live post that stayed in network has no published cooldown. The next post is a new candidate. These sources do not publish a 14-day recovery.

Three posts after the sheet

The sheet is the diagnosis. The next three posts are the test. Hold every variable you did not name in the change column.

Write the receipt before you open a model. One incident from this week, one refusal, one mechanism, one line a stranger can use. If you cannot fill those four lines, you do not have a post yet. You have a topic. The reason the sentence has to come from material you brought is the constraint in building Boykli.
Draft from that note. On Boykli, Create posts takes the note, the link, the file, or the Voice Agent transcript. Short drafts sit in a 45–400 character band, medium in 400–1400, long in 1400–3000. Pick the band Window B already used, so length is not the new variable. Edit until you would say the paragraph on a call.
Read the line breaks before you approve. The preview shows the draft at phone width and desktop width. The first lines are what a member sees before "see more." The preview does not predict impressions. How that check works is the post preview page.
Queue one approved item to the personal profile. The Schedule page stores the weekdays you toggle, one time of day, and one timezone. It sends one due item at a time. It does not watch the first hour, leave a first comment, or move the slot because yesterday's median was low. Whether a scheduled send and a hand publish face the same content rules is scheduled posts and reach.
Stay for replies you can add a fact to, if you are at the screen. Five comments you type are a different act from a tool that comments for you. The March 2026 notes treat automated comments as inauthentic. They are not a reach repair.
The next day, copy the same four analytics fields onto the sheet. After the third post, take the median of those three and set it next to Window B. Three posts is a small sample. Call a move only if the median and the in-network share moved in the direction your pattern predicted. If they did not, you learned that this change was not the lever. You did not learn that the account is flagged.
Only then consider a second change. Format is the usual second test, and it needs its own three posts with the text class held still. A carousel does not restore a generic paragraph. Dwell is format-relative. The engagement-rate claims around document posts are separated from LinkedIn's own notes in carousel engagement.

Boykli can hold steps 2 through 4. Managed caps are 210 posts, 150 carousels, 150 images, 25 Voice Agent sessions, and 130 post refinements a month. BYOK uses your OpenRouter key for chat and images and your Google AI Studio key for Voice Agent, with no Boykli generation cap. You pay those providers for usage. Neither seat imports LinkedIn analytics. If enough "Seems like AI slop" reports cross LinkedIn's unpublished threshold, the September note says the tip appears in LinkedIn's post analytics, not in Boykli.

FAQ

How do I know my LinkedIn reach is actually declining?

Compare the median impressions of at least eight posts in a recent window with the median of at least eight earlier posts of the same format. Ignore anything newer than yesterday. If that median is down, call it a decline, then read members reached and the in-network share. If the only baseline is your best post, you do not have a decline yet. You have an outlier.

Is declining reach a shadowban?

LinkedIn has not published a shadowban flag or a penalty score. A restriction is a notice with a time, and it follows automated activity on the website. A removed post is gone under the community policies. A live post whose displays stayed with people who already follow you is the distribution taper described in the June and September 2026 notes. Those are three different incidents. The median sheet tells you which one you have.

Will posting more often fix a reach decline?

Posting more changes the monthly total and gives you a larger sample. It is not a published way to raise the per-post median. Repeat generic posts are the class the June and September notes say are less likely to travel past the immediate network. During the three-post test, keep the weekly count where the recent window already was.

Does changing the posting time restore reach?

Not as the first move. Change the text class that matches the pattern you measured, and keep one clock so the next median is readable. The first hour can move a popularity percentile within minutes. It is not a published on-off switch at minute 60, and it does not explain eight posts of the same format sitting below the earlier window.

Should I delete old posts to reset the algorithm?

These sources do not describe a reset from deleting posts. A removed post stays removed. A restricted feature returns at the time on the notice, after the software is disabled. The next post you publish is a new candidate. Judge it with the same four fields: impressions, members reached, in-network percent, and out-of-network percent.

Do carousels fix declining reach?

A document post is a different format, with format-relative dwell. It can be the second experiment, after the text class is clean and the format is the only thing you change. It does not override a bait closer, a pod, or a paragraph any account could sign. Boykli can build the PDF from a note and a brand kit. It does not attach an impression multiplier.

Can Boykli diagnose my LinkedIn analytics?

No. The Algorithm page explains the public notes. Library stores what you wrote in Boykli. The queue sends the approved item. Your discovery block stays on LinkedIn. You copy impressions, members reached, and the network split yourself.

What does a seat cost if I only want the draft and the queue?

BYOK is $29 a month. Managed is $88 a month. Browsing Radar and Algorithm does not require a plan. Writing, Voice Agent, publishing, and scheduling do. The seat does not buy distribution. An AppSumo lifetime code redeemed at /redeem is a BYOK seat without a Dodo subscription. One code, one account. A revoked code cannot be reused.

Diagnose on LinkedIn, then draft from the note

Approve a post you can sign, then queue it on your personal profile. BYOK is $29/month. Managed is $88/month. Boykli will not score the decline or run a recovery pod.