DEJAN launched an Ablation tab in its Content Optimizer on 20 October 2026 to measure how many rank places each sentence in an AI-cited source is worth. (Ablation: A method of measuring the value of specific components by testing combinations of their presence or absence.) The tool identifies specific “spans” of text that drive ranking, such as a +4.10 ± 0.10 rank place gain for a single sentence in a test case. This impacts GEO practitioners who need to optimize content for AI-grounded results. (GEO: Generative Engine Optimization which involves optimizing content for results grounded in artificial intelligence.)
Is this for you?
GEO practitioners can use the Ablation tool to identify the 20% of text that drives 80% of their ranking gain.
- If you are a GEO specialist, the tool identifies which specific spans of text drive your ranking gains.
- If you are a technical SEO engineer, you can empirically test the weight of signals like brand names versus factual content.
- If you are a content strategy lead, you can prioritize high-value spans to improve AI visibility.
- Skip this if you are a beginner blogger as it requires an understanding of Shapley values and AI grounding mechanics.
What you can do now
- Set a cost cap in the tool to prevent the span test from exceeding a specific budget.
Key figures
- Ablation availability
- Open to all users as of 20 October 2026
- Identified sentence value
- +4.10 ± 0.10 rank places
- Brand name influence
- +0.10 ± 0.10 (effectively zero)
- Value concentration
- 20% of text holds 80% of the gain
- Test execution cost
- $3.02 for 188 rankings
The Shapley value methodology for rank attribution
Ablation uses Shapley values—the average number of rank places a target gains when a specific span is present across many combinations—to provide precise attribution. (span: A discrete unit of text like a sentence or paragraph used as a unit of measurement.) The tool first identifies the cited source, cuts it into spans (discrete units of text like sentences or paragraphs), and ranks the target many times with different combinations of these spans removed. This method provides higher accuracy than older techniques because, in DEJAN’s testing, removing a single signal at a time failed to move the rank at all.
The system calculates the value of each span by measuring its impact
The system calculates the value of each span by measuring its impact across these multiple combinations. If a span’s value range includes zero, it indicates the span has no measured effect on the ranking. This mathematical approach allows the tool to isolate the specific contribution of a sentence from the surrounding context.
Case study: Analyzing the Semrush Sensor results
The Semrush Sensor test demonstrates how specific information density drives visibility. In this run, the target source was grounded and cited first by Google, with a baseline rank of 3rd. When the full text was removed, the rank dropped to 8th, revealing a 5-place difference that the tool then decomposed.
One specific sentence regarding monitoring 20+ categories
One specific sentence regarding monitoring 20+ categories was found to be worth +4.10 ± 0.10 of those 5 places. Conversely, the brand name “Semrush” appeared four times but was tested as a single span, yielding a value of +0.10 ± 0.10. Because this range includes zero, the brand name has no measured effect on the ranking.
Value concentration in the text
Value concentration in the text is remarkably high. DEJAN reported that 20% of the text holds 80% of the gain, showing that a small percentage of content drives the majority of results. The test was completed using 188 rankings at a cost of $3.02.
Brand mentions vs. information density
Information density is significantly more impactful than brand mentions in AI citations. In the Semrush example, the brand name provided no measurable rank gain while a single fact-heavy sentence provided nearly the entire gain. This suggests that for Generative Engine Optimization (GEO), the presence of a brand name is a secondary signal compared to the specific data provided.
In AI citations
Identifying the small percentage of text that drives the most results allows for extreme content efficiency. By isolating the 20% of text that holds 80% of the gain, editors can prioritize high-value spans. In AI citations, providing useful information is significantly more impactful than simply mentioning a brand name.
Technical configurations and platform support
The Ablation feature supports Google, OpenAI, and Claude, with different data handling for each. Google uses the real snippet text the answer was grounded on or a synthetic snippet; OpenAI uses the real snippet text the model read; and Claude uses a synthetic extractive snippet because it does not return snippet text. (synthetic snippet: A piece of text generated by a model when the original source does not return a specific snippet.) Users can choose between Snippet mode, which ranks the snippets, and Page mode, which fetches and ranks the full pages.
Users can adjust the span unit depending on the mode, choosing sentences for Snippets or paragraphs for Pages. A cost cap can be set to prevent the span test from exceeding a specific budget, as the tool ranks many combinations to determine the final Shapley values.
Who this is for: GEO (Generative Engine Optimization) specialists
GEO (Generative Engine Optimization) specialists should use this to identify the 20% of text that drives 80% of their ranking gain. Technical SEO engineers can use the tool to empirically test the weight of signals like brand names versus factual content. Content strategy leads should use the tool to focus on high-value spans rather than just keyword inclusion to improve AI visibility.
Who this is not for
Beginner bloggers should skip this tool as it requires an understanding of Shapley values and AI grounding mechanics. These readers should instead look for general guides on writing for AI search engines.
A scenario for evaluating content value
Yusuf runs a tech blog and wants to rank for “mobile vs desktop search volatility.” He uses the DEJAN Content Optimizer to run an Ablation test on his target source. The tool reveals that one specific sentence about tracking 20+ categories provides a +4.10 rank gain, while the mention of his own brand provides a +0.10 gain (essentially zero).
If Yusuf were to remove the high-value sentence about the 20+ categories, his page would drop from 3rd to 8th place in the AI-grounded results. However, if he removed only the brand name, his rank would likely stay at 3rd because the brand name has no measured effect.
What this is used for: The Ablation feature offers several practical applications for content
The Ablation feature offers several practical applications for content optimization:
- Identifying the “power law” of content (20% text / 80% gain) $ ightarrow$ Content Editors
- Benchmarking the rank impact of specific sentences $ ightarrow$ SEO Practitioners
- Comparing snippet vs. full page ranking value $ ightarrow$ Technical SEOs
- Validating if brand mentions are necessary in body copy $ ightarrow$ Content Strategists
What this replaces: Ablation replaces manual signal testing
Ablation replaces manual signal testing, where a user removes one signal at a time to see if the rank moves. The new method is more accurate because, in DEJAN’s tests, removing one signal at a time did not move the rank, whereas measuring many combinations did.
What is not yet known: DEJAN did not specify the exact hardware or software requirements
DEJAN did not specify the exact hardware or software requirements for running the Ablation tool. The company also did not disclose known problems or limitations of the Shapley value calculation in specific edge cases. (Shapley value: The average number of rank places a target gains when a specific span is present across many combinations.) Further results are needed to confirm if these concentration patterns hold true across different industries beyond the tech sector.
Ablation: what is known and what is not yet disclosed
- Who gets it, and when — which devices, plans or regions?
- The Ablation tab is open to all users. The original report did not say which devices, plans or regions it is available in.
- What does it require, and is anything removed or changed for existing users?
- The original report did not say what it requires or what was removed for existing users.
- Are there known problems or limitations?
- The original report did not say what the known problems or limitations are.
Key terms
- Ablation
- A method of measuring the value of specific components (spans) by testing combinations of their presence or absence.
- Shapley value
- The average number of rank places a target gains when a specific span is present across many combinations.
- Span
- A discrete unit of text, such as a sentence or paragraph, used as the unit of measurement in an ablation test.
- Grounded prompt
- A prompt that forces an AI model to use specific cited sources to generate an answer.
The ability to quantify the exact weight of a sentence moves SEO from guesswork to precise engineering. By identifying which parts of a content piece are actually responsible for visibility, growth engineers can optimize for the highest possible ROI on every word.
Source story: New Feature: Grounding Snippet Ablation, published by DEJAN on 20 October 2026.