We’ve been link building for over two decades now and have developed a link integration philosophy aimed at effectiveness and safety for our clients.
Naturally integrated, undetectable links.
The advantage of our approach to links is in our strict link integration guidelines. We aim for low-to-no detectability. This means that if someone was to take a look at a page where you’re linked from they won’t be able to tell who wanted the link there and if they do, that link will be perfectly justified by purpose and context.
We don’t subscribe to the quantity and authority method.
If you’re after an exact number of links each month and they all need to be of a certain arbitrary metric provided by your favourite tool, then we’re probably not a good match. If you’re willing to align with our philosophy of natural link integration and undetectable links above all else then we should talk.
Our early research into links on the web dating from 2011 to 2016 shows that the majority of link builders and SEOs simply don’t know how to integrate links well. A trained eye can always spot “the one who wanted the link” and most efforts to make links “look natural” are fruitless.
A 2023 review of the state of the links on the web discovered that absolutely nothing has changed and, in fact, some of the industry’s less impactful rules have also become commonplace in the blogosphere.
Even when a good link is made, it’s generally devalued by less than ideal integration.
At SMX Munich in 2023 we presented our statistical model of links on the web based on the most common keywords found in natural links on the web and shared the tool with the audience.
Building on this work later that year we started gathering data and planning development of a machine learning model that understands the links on the web intuitively and can generalise well enough to be deployed on a wide variety of link related tasks and workflows.
In early 2024, we successfully trained a general-purpose transformer model and announced a public release of its smaller variant LinkBERT on HuggingFace.
LinkBERT is a fine-tuned version of BERT (large, cased) trained on a 4.5GB dataset on top quality content and links on the web. This transformer is designed for binary token classification. When presented with plain text it can predict which parts of that text are most likely to be links / anchor text.
LinkBERT’s applications are vast and diverse, tailored to enhance both the efficiency and quality of web content creation and analysis:
A public demo of the model is available at: https://dejan.ai/linkbert/
