Definition
Position-adjusted word count (PAWC) is a visibility metric for AI-generated answers that measures how many of an answer's words are attributed to a given source, giving more weight to sentences that appear earlier in the answer. It was introduced in the paper that coined the term generative engine optimization, GEO: Generative Engine Optimization by Pranjal Aggarwal (IIT Delhi), Vishvak Murahari, Ameet Deshpande and Karthik Narasimhan (Princeton) and two independent co-authors, first posted in November 2023 and published at KDD 2024 in Barcelona.
The authors needed a new metric because ranking does not translate to GEO. Classic SEO measures visibility by average position in a list of links. A generative engine blends several sources into one response and attributes them inline, with different lengths and at different points in the text, so a source can be cited once in passing or carry half the answer.
PAWC is built in two steps. The plain word count metric divides the number of words in sentences that cite a source by the total words in the response; when a sentence cites several sources, its words are shared equally between them. PAWC then multiplies each sentence's word count by a weight that decays exponentially with the sentence's position, so words near the top count more than words near the end. The authors justify the decay by pointing to search studies showing click-through rates fall off steeply with rank. As a result, a source cited in the opening sentence can outscore one with more cited words lower down.
The paper pairs PAWC with a second metric, Subjective Impression, scored by an LLM judge (G-Eval) on seven facets: relevance of the cited material, influence on the answer, uniqueness, subjective position, subjective count, likelihood of a click and diversity of material. PAWC is the objective half of the pair.
The metric produced the paper's headline results. On GEO-bench, a 10,000-query benchmark answered by GPT-3.5-turbo from the top five Google results, three content changes improved PAWC by 30% to 40%: adding quotations, adding statistics and citing sources. The best methods beat the baseline by 41% on PAWC and 28% on Subjective Impression. Keyword stuffing scored below the unmodified baseline. Gains were largest for low-ranked sources: citing sources raised the visibility of the fifth-ranked result by 115.1% while the top-ranked result lost 30.3% on average. On a 200-query test with Perplexity, quotation addition improved PAWC by 22% and keyword stuffing did 10% worse than baseline.
PAWC is hard to reproduce on commercial engines. It needs sentence-level attribution, while ChatGPT, Google AI Overviews and others often attach sources to a paragraph or the whole answer, and their answers vary between runs. Most AI visibility tools therefore use simpler or broader measures: mention rate, answer position, citation rank for the order of sources, and composite scores. Promptwatch's AI visibility score, for example, combines a brand's placement in the answer, its share of attention and the depth of coverage, which follows the same idea that prominence, not just presence, determines visibility. PAWC remains a useful reference when comparing GEO performance metrics and when designing controlled tests of LLM citations.
Examples of Position-Adjusted Word Count (PAWC)
- In a five-sentence AI answer, source A is cited only in the first sentence while source B is cited in sentences four and five with twice as many words. On plain word count B wins clearly; on PAWC the gap narrows because A's words carry a higher position weight.
- A research team rebuilds the GEO-bench setup with its own retrieval pipeline to test whether adding customer statistics to product pages raises their PAWC, before rolling the change out across the site.
- An analyst adapts the idea to brand tracking by weighting each sentence that names the brand by its position in the response, producing a prominence score that favors being the first recommendation over a mention in the closing paragraph.
- In the GEO paper's experiments, rewriting a source with more query keywords lowered its PAWC from the baseline of 19.3 to 17.7, one of the clearest signs that keyword stuffing does not carry over from SEO.
