An AI writing tool will rewrite your opener forty ways in nine seconds. None of the forty know that the person read your competitor's post on Tuesday, or that they are one of eleven people in your whole category who did.

That is the gap. Personalization software has been solving the writing problem for four years. The writing was never the problem. Knowing who is on the other end was.

This piece covers what AI personalization tools actually do, where each category stops, and what changes when the input is a named market rather than a scraped profile.

What AI personalization tools actually do on LinkedIn

There are four categories on the market and they do different jobs. We are describing categories rather than naming products, because tool names churn every quarter and the categories do not.

Profile summarisers. They read a LinkedIn profile and produce a line about the person: their role, their last job change, a skill they list. The output is fluent and usually true. It is also available to every competitor who runs the same tool on the same profile, which is why so many first lines now sound identical.

Post-comment openers. They read the person's recent posts or comments and write a first line referencing one. Better, because it references something the person chose to do rather than something written on their profile three years ago. The limit is coverage. Most people in most B2B markets post nothing. If your target has not posted in eight months, this category has nothing to work with.

Sequencer and CRM writing assistants. They sit inside the tool you already send from and rewrite subject lines, adjust tone, and generate variants. Useful for volume and consistency. They have no information about the person that your CRM did not already hold.

Research agents. They browse, read the company site, find a funding round or a job ad, and write a paragraph on why now. This is the strongest category and the slowest. It works well for twenty accounts a week and falls apart at four hundred.

All four take the person as given. You supply a name, they supply the words. Nothing in that loop tells you whether the name was worth writing to.

Why AI personalization fails when you do not know your market

LinkedIn has more than 1 billion members in more than 200 countries and territories, as published on about.linkedin.com, September 2026. A job title filter over that is a very coarse instrument. "Head of Customer Experience" in a European ecommerce market returns thousands of people, most of whom have never looked at a vendor in your category and never will.

Run the best writing tool in the world over that list and you have written thousands of well-crafted messages to people with no interest in the problem. The reply rate tells you the copy failed. The copy did not fail.

Here is what the difference looks like with numbers. In an ecommerce customer-service AI market pulled in London in August 2026, 264 people sampled and the 170 most active analysed, the vendors ranked like this on market reactions per post: Kustomer 6.6, Gladly 3.2, Siena 3.1, Yuma 2.2, Loop Returns 2.0, Ada 1.3, parcelLab 1.2, Zowie 0.9, Gorgias 0.8, Narvar 0.2. Gorgias is the biggest ecommerce helpdesk brand in that set and it is close to invisible in that market's attention.

If you sell into that market, the useful fact is not "personalise your outreach". It is that 170 specific named people are the ones paying attention, and that they react to Kustomer at eight times the rate they react to Gorgias. That changes who you write to, what you reference, and what you assume they already know.

The input that makes a message personal: a named market

A named market is a list of the actual people who decide and influence purchases in one category, with name, role and LinkedIn profile, kept current. It is built from six things: the people who react to and comment on your competitors' posts, the people who engage your own page and team, the creators that market follows, past customers who moved company, rising topics in the category, and your own customer list.

Two properties matter for personalization.

It is small enough to be honest. A market of 2,500 named people is not a sample of a large population. It is close to the population. When you write to someone in it, you know why they are in it.

It carries provenance. Every person arrived through a route you can name. This one reacted to a competitor's post about onboarding. That one has commented on the same creator three times. The other one used your product at a company they left in March. A first line built on that is checkable, which is a different thing from plausible.

Provenance is what most personalization tools are missing. They generate a reason to write. They do not have one.

Captured engagement: the post, the reaction, the profile view

Once the market exists, capture keeps it alive. Every reaction, comment and profile view from those people comes back as a named person with the exact post attached. Profile views need Premium or Sales Navigator on the connected seat.

That gives you three honest openers that no writing tool can invent.

The person reacted to a specific post. Not "I saw you are interested in customer experience". You know the post, the date and what it said. In one revenue-tooling market in the Nordics, pulled 27 August 2026 over a 90-day window with 170 of 642 sampled people analysed, Gong's best post was "the biggest sales call mistake" with 74 market reactions. Eleven of your buyers reacting to one named argument is a real thing to open with.

The person viewed your profile. You did not guess at intent. They came to you.

The person moved. Someone who used your product at a previous company and is now somewhere else is a warm start with a fact attached.

None of this requires a scraper, and it should not use one. LinkedIn's User Agreement prohibits developing, supporting or using "software, devices, scripts, robots or any other means or processes (such as crawlers, browser plugins and add-ons or any other technology) to scrape or copy the Services, including profiles and other data from the Services", as published in LinkedIn's User Agreement, September 2026. Public engagement on a connected seat is a different route to a different kind of fact.

How to test an AI personalization tool before you pay for it

Five questions, and the answers are usually quick.

  • Where does it get the person? If the answer is "you upload a list", the tool is a writer, not a research product. Price it as a writer.
  • Can it tell you why this person and not the next one? A tool that cannot rank people cannot tell you who to skip, and skipping is most of the work.
  • What happens when the target has not posted? Ask for the fallback. If the fallback is a profile summary, you are back to the line everyone else sends.
  • Does anything update weekly? People move company. Attention shifts between vendors. A static list decays fast.
  • Would you send the output to yourself? Read three generated openers out loud. If the first clause could be sent to any of forty people, it will be.

Across four markets we have pulled so far, craft content beats product news by about two to one, and founders and operators out-reach their company pages by roughly ten times. Both of those change what a personal message should reference. Neither is something a writing tool can find out on its own.

What we do and what we do not do

Postelix builds the named market and captures the engagement. Draft testing runs against the market before you post, so you know which angle the market reacted to rather than guessing. Follow-up drafting is included from Tier 1 at EUR 490 a month.

What we do not do: we do not promise reply rates, we do not push into a sequencer for you, and we do not claim a message will land because a machine wrote it. The drafting is the easy half. The list is the half that decides whether the message was worth writing.

Questions

What is the best AI tool for LinkedIn personalization?

There is no single best one, and the category you need depends on volume. For twenty accounts a week, a research agent that reads the company site produces the strongest openers. For four hundred, only a maintained list of people with a known reason for being on it will hold up. The writing tool is the smaller decision.

Does AI personalization actually improve reply rates?

It improves the copy. Whether that improves replies depends on whether the list was right in the first place. A well-written message to someone who has never looked at your category still fails, and the failure looks like a copy problem when it is a targeting problem.

Is scraping LinkedIn profiles for personalization allowed?

LinkedIn's User Agreement prohibits using software, scripts, robots, crawlers or browser plug-ins to scrape or copy data from the service, as published in LinkedIn's User Agreement, September 2026. Working from public engagement on a connected seat, and from your own customer data, is a different route.

How do you know someone reacted to a competitor's post?

Reactions and comments on public posts are public. We collect them for the people in your market and keep the exact post attached, so the reference in a message is checkable. Profile views are separate and need Premium or Sales Navigator on the connected seat.

How big should a named market be?

Basic costs EUR 490 a month and discovers up to 2,000 new people per month, with capacity for 6,000 active people in one market. The markets we have pulled have run from a few hundred highly active people to several thousand named ones. Smaller and specific beats large and vague every time.