AI filler detector

Why we made this tool

There is a silent problem in much of digital content: a huge proportion of the texts published as "professional writing" are actually AI-generated content with minimal editing. Fillers are the most visible clues. Phrases like "in today's digital world", "it's not just X it's Y", "whether you're", "tailored solutions", "cutting-edge" appear over and over in agency websites, corporate blogs and service pages. They are not there by mistake: they are the typical tics of certain language models when asked to "write a commercial text about X".

The problem for the client or the reader is that those texts communicate little, generate gradual distrust and, most importantly for us, lose ranking. Google has penalized low-quality auto-generated content since 2022 with its Helpful Content updates. New-generation AI engines like ChatGPT, Perplexity and Gemini discard as a reliable source content they detect as AI-generated. The paradox is notable: today writing badly with AI loses on both fronts at once, in classic SEO and in AEO.

This tool is the simplified version of the internal filter we apply to any text before publishing on this site. It is not a magic wand or a perfect AI detector: no detector is, not even paid ones. But it is a useful and honest first filter that anyone can use to measure their own content, evaluate what they receive from providers, or audit an agency's blog before hiring them.

What it detects and what it does not

The tool looks for four categories of signals. First, specific fillers: a base of around 30 set phrases that appear with disproportionate frequency in auto-generated texts ("in today's digital world", "unlock the power of", "dive into", "cutting-edge", "tailored solutions", "it's not just X it's Y", "whether you're", "it's important to note", "when it comes to"). Second, structural patterns: filler tricolon (lists of three adjectives without substance, like "fast, efficient and effective"), overly clean symmetric phrases, hollow transitions ("in this article we'll explore", "as we've mentioned", "in summary"). Third, artificial repetition: the same syntactic structure repeated in consecutive phrases, characteristic of automated generation. Fourth, linguistic emptiness: adjectives without substance ("dynamic", "innovative", "seamless", "robust") when not followed by a concrete explanation.

The tool does NOT detect three things worth mentioning so as not to create false expectations. It does not identify whether a text was specifically generated by GPT-4, Claude or Gemini —no detector does so reliably—. It does not evaluate the factual accuracy or the truthfulness of the content —the text may sound impeccable and state incorrect data—. And it does not judge the quality of the argument or the analytical depth: a text can pass the filler filter and still be superficial in its content. The tool covers one concrete axis of the problem, not all the axes.

How to interpret the result

The score the tool delivers goes from 0 to 100, where 100 is a solid text with no AI signals and 0 is a text riddled with automated-writing patterns. The practical interpretation of the ranges is as follows. Between 85 and 100, the text is well written and requires no intervention on account of fillers; it may have other problems but not this one. Between 65 and 85 there is room for improvement: the text is acceptable but contains signals worth reviewing and rewriting before publishing. Between 45 and 65 there are clear signs of mechanical or assisted writing without serious review; the text needs significant rewriting. Below 45 the text sounds strongly like auto-generated content and requires a complete rewrite by a person, not superficial editing.

More useful than the final number is the list of specific signals detected. Each filler or pattern appears with its context: the exact phrase and a couple of words around it. That allows going directly to the points to correct without having to reread the whole text looking for them. A good practice when rewriting is not to delete the phrase as is: instead, ask yourself what that phrase really wanted to say and reformulate with concrete words. If the answer is "I don't know what it wanted to say", the phrase was filler and is deleted without losing anything.

Privacy and technical operation

The entire analysis happens in your browser. The text you paste does not leave your device: it is not sent to any server, not saved in databases, not used to train models. When you close the tab, everything disappears. This technical decision is deliberate and allows anyone to use the tool with confidential content, drafts, unpublished proposals or any other sensitive material with no risk of leakage.

The operation is simple: we load a base of patterns (fillers, structures, set phrases) on the client and apply local text analysis with JavaScript. The initial load is a few kilobytes, so the tool works even on slow connections and modest devices. The pattern base is updated when we detect new tics characteristic of recent language models; the updates are transparent and require no user intervention.

Typical uses of the tool

Three usage scenarios cover 90% of the cases we see. The first is self-evaluation: a company that writes its own content pastes the texts before publishing and adjusts them until they cross the threshold of 85. The second is provider quality control: a company that hires external writing pastes what it receives to verify it is real human work and not content generated with minimal editing. The third is competitor auditing: pasting content from a competitor or from an agency you are evaluating to hire to see what kind of writing it produces.

For all three uses, the tool is a starting point, not a final verdict. A text with many fillers may have been written by a person with bad editorial habits; a text without fillers may have been generated by a well-tuned model. What a low score does indicate is that the text, regardless of its origin, sounds like AI, and that already has consequences in SEO and AEO even if it was written by hand. The operational question is not "was it written by AI?" but "does it pass the filter of a qualified reader and of modern algorithms?". When the answer is no, it is worth rewriting.

The methodology behind the score: how it is calculated

For the tool to be more useful than a black box, it is worth explaining how the final score is calculated. We start at 100 and subtract points for each signal detected, weighted by its weight. The most characteristic fillers of auto-generated writing have weight three (those that appear in any typical AI-generated text), the hollow transitions and filler tricolons have weight two, and the mild lexical emptiness has weight one. The accumulated total is normalized by the length of the text so as not to unfairly punish short texts or give a free pass to long texts with many signals.

The normalization is important. A 100-word text with three fillers has a higher problem density than a 3,000-word text with the same three fillers. The score reflects that difference: the first drops more than the second, even though both have the same absolute number of signals. In practice this means well-written long texts can have a couple of isolated fillers without the score collapsing, while short texts riddled with patterns stay clearly below the acceptable threshold.

Honest limitations: what this tool does not solve

It would be dishonest to close without talking about the limits. This tool covers a real axis of the text quality problem, but there are other axes it does not touch that matter as much or more. The first is factual accuracy: a text can pass the filler filter and confidently state false data. To verify truthfulness, human research or specialized fact-checking tools are needed. The second is argumentative depth: a text can be free of fillers and be superficial, repeat the obvious or contribute no genuine analysis. Fluency does not equal substance. The third is editorial structure: a text well written phrase by phrase can be badly organized, with paragraphs that do not advance, disconnected sections or an absence of logical flow.

That is why we think of this tool as a first filter, not a final verdict. A text that crosses the threshold of 85 has passed an important test but still needs critical human reading before being published. A text below 65 has a visible problem that must be solved before even evaluating other aspects. The tool separates the noise from potentially useful content, but the final evaluation of usefulness still requires a qualified human eye. The earlier in the publication flow this filter is applied, the less time is wasted editing texts that did not deserve editing but a complete rewrite.

How the problem evolved since 2022

It is worth situating the phenomenon. Before the end of 2022, auto-generated content was marginal: it was produced by simple generators that sounded clumsy, and any skilled editor detected them in seconds. The massive irruption of high-quality language models changed the equation in a matter of months. Suddenly anyone could ask for a "professional article about web design" and receive a thousand coherent, grammatically correct and completely empty words. The technical barrier to producing content dropped to zero.

The sector reacted in two opposite directions. On one hand, agencies and individuals who saw an opportunity to scale production: publishing ten times more articles, maintaining inflated editorial calendars, filling websites with endless sections, all at a marginal cost close to zero. On the other, platforms and readers who learned to distrust. Google adjusted its algorithms with several Helpful Content updates between 2022 and 2024 that specifically punished the mass production of low quality. Qualified readers developed a sharpened sense: when a text begins with a recognizable cliché, half close the tab before the second sentence. The mass production of auto-generated content revealed itself as a shortcut that ages badly and that in 2026 is starting to be actively disadvantageous.

The detector you just used was born in that context. The fillers it identifies are patterns that appeared, amplified and became clichés over the last four years. The pattern base is updated when we detect new lexical fads generated by recent models; the patterns of 2023 are not exactly the same as those of 2026.

The signals most seen by type of content

The distribution of fillers varies by the type of text. Commercial texts for service websites —"about us" pages, service lists, proposals— concentrate the aspirational fillers: strong online presence, unique experiences, attractive and functional designs, solid identities. Marketing blogs repeat the hollow transitions and filler tricolons. AI-generated technical texts tend toward impersonal constructions and excessive enumerations, with phrases that start with an empty transition and end with general verbs like optimize, enhance or maximize with no concrete object.

Identifying your own pattern helps correct it. If your detector flags a lot of lexical emptiness, the text is probably a commercial presentation without concrete content: the correction consists of replacing each empty adjective with the specific data or example it was meant to summarize. If it flags many hollow transitions, the text was probably generated in a single pass without later editing: it is worth doing a second pass removing the filler connectors. If it flags filler tricolons, it is a sign that the generator or the writer was stretching length without substance; replacing the empty lists with a single precise word almost always improves the text.

Why it matters specifically for AEO

This tool has an indirect but critical usefulness for AI engine optimization. Models like ChatGPT, Perplexity and Gemini do not cite as a reliable source content they detect as AI-generated, and the patterns they recognize are precisely the ones this tool identifies. Passing the filler filter is a necessary condition to have a chance of being cited by generative engines.

The logic from the model's side is direct. If an LLM cites as a source a text that another LLM generated, the resulting answer becomes circular: information with no added value, empty of original data, that contributes nothing to the user asking. AI models are trained to avoid that circularity because it destroys their usefulness. The practical result is that sites full of auto-generated content become invisible to AI engines, even when they have acceptable traditional SEO authority. The low-text-quality signal erodes citability.

That is why we recommend using this tool at two moments in the publication flow. Before publishing: review any text going to the site, especially if it went through a generator or a writer with no time for serious editing. And periodically on old content: paste articles published months or years ago to audit whether they sound like their writing era and need a refresh. Small edits removing fillers can recover citability in content that was aging badly.

And to close with the operational question many users ask: how much work does it cost to cross the threshold once the signals are identified? The practical answer is: much less than it seems. Most of the texts we detect below 65 can be brought above 85 with two or three review passes, removing the most obvious fillers and rewriting the empty phrases with concrete words. The effort is modest compared to the benefit: a text that goes from 60 to 90 improves simultaneously in SEO, in AEO and in reader trust, without having to add new content or redo the structure. It is one of the corrections with the best cost-benefit ratio in digital writing, and the tool is here to speed it up.

One last practical observation: if a text crosses the threshold of 85 but your editorial intuition keeps telling you something is not right, listen to your intuition before the score. The tool detects visible lexical patterns, it does not detect subtleties like weak argumentation, lack of concrete examples or overuse of the passive voice. Those problems require a qualified human eye and are precisely what separates a correct text from a memorable one. The detector is good infrastructure, not good editorial judgment, and it is worth treating it as a useful assistant subordinate to the head of whoever writes the text, not as a substitute for the craft that no algorithm fully replaces in the real practice of careful editorial work, especially when the texts carry commercial or reputational weight for the brand that publishes them.