Dan Petrovic ·
20 October 2026
The new Ablation tab in DEJAN's Content Optimizer measures how many rank places each sentence of an AI-cited source is worth.
Content Optimizer in the DEJAN app has a new Ablation tab, open to all users. It runs a grounded prompt with Google, OpenAI or Anthropic models, takes the sources the AI answer cited, and measures how many rank places each sentence or chunk of the top cited source is worth. In our public example, one sentence of the source Google cited first is worth 4.1 of the 5 rank places its text earns, and its brand name has no measured effect.
We ran the prompt "Are there trackers that compare desktop versus mobile Google search volatility?" on Google from Australia, in our public demo project. Every figure in this post is a screenshot of run 33, where you can check each number.
Cited sources (Sources tab). Each line follows one source from the order Google grounded on, to the order it cited them in the answer, to the baseline rank our ranker gave it. The target, highlighted in blue, is grounded and cited 1st and ranked 3rd.
Google cited eight sources. The target is the one it cited first, Semrush's Sensor page. Our ranker placed it 3rd in all three baseline rankings, behind flyingvgroup.com and martech.zone.
With its text removed the target ranks 8th, and with its full text it ranks 3rd. One sentence is worth +4.10 ± 0.10 of those 5 places:
Semrush Sensor measures volatility in search results, tracking down 20+ categories on mobile and desktop and highlighting possible Google Updates.
The brand name, Semrush, appears four times and is tested as one span. It is worth +0.10 ± 0.10, which is inside its range of zero. The span that pulls the target down most is "You can use tools like Semrush's Traffic Analytics to guide your marketing strategies based on what works for your competitors." at −0.20 ± 0.13 places. The test took 188 rankings and the run cost $3.02.
The run page shows the result in seven views. Each figure below is one of them.
Rank places each span is worth (Main tab). Each bar is one span's Shapley value with its ± range. Green bars hold the rank up and red bars pull it down. A bar whose range reaches zero has no measured effect.
Heat map (Text tab). The target's own text, coloured by value. The darker the green, the more rank places a sentence holds; red sentences pull the rank down. Brand mentions are underlined with dots and tested together.
Value map (Text tab). Each brick is one span, as wide as its word count, in text order. The dashed brick is the brand. Here the first brick holds most of the value.
How concentrated the value is (Main tab). The share of the rank gain against the share of the text, with spans taken from most to least valuable. The dashed line is an even spread. In run 33, 20% of the text holds 80% of the gain.
Rank by how much text is kept (Main tab). The target's average rank across rankings that kept the same number of spans, from no text on the left to the full text on the right. It rises from 8th to about 3rd.
How often each rank came up (Main tab). The share of all 188 rankings that placed the target at each rank. Two peaks, 3rd at 47% and 7th at 37%, and almost nothing between them.
Ranking stream (Ranking stream tab). The latest 60 rankings, newest on top. Each row is one ranking and each column one span in text order, with the brand last. Empty squares are removed spans, and the label is the rank that ranking gave the target.
The two peaks come from the one sentence. In the 60 rankings in the stream, all 33 that kept it placed the target 2nd or 3rd, and all 27 without it placed the target 7th or 8th.
Step 1 runs the prompt, collects the text of each cited source, and ranks the unchanged set three times with the Content Optimizer ranker. The median is the baseline rank. Step 1 then cuts the target into spans. A span test typically costs between $0.50 and $20, depending on the size of the text, and never more than the cost cap. The default target is the source the platform cited first, and you can pick another.
Step 2 is the span test. It ranks the set many times with different combinations of the target's spans removed, and gives each span its Shapley value: the average number of rank places the target gains when that span is present. When the source names its brand, every mention is tested together as one span. The test stops when every ± range is 0.25 places or less, at the cost cap, or when every combination has been ranked.
We made the span test the default because in our tests, removing one thing at a time (a signal, the brand, a single sentence) did not move the rank. Measuring each span across many combinations did.
The method is called CHISEL, short for Chunk Holdout Influence Scoring by Elimination Loss. The paper, with results from 111 runs on Google, ChatGPT and Claude, is at dejan.ai/media/pdf/CHISEL.pdf.
Open a project in the DEJAN app and go to Content Optimizer → Ablation. Past runs are listed under the form. On pay-as-you-go projects both steps are billed from the credit balance, and step 2 starts only when the balance covers its cost cap.
