/terms/pawc · 4 min read · advanced

Position-Adjusted Word Count

Position-Adjusted Word Count (PAWC) is the metric in Aggarwal et al. 2023's GEO paper that scores how much of an AI engine's answer is drawn from a given source, weighting earlier-positioned text more heavily. It is the number behind nearly every '+40% GEO visibility' claim, but it measures word-count share under single-actor 2023 conditions, not citation rate or ranking.

Citation status

ChatGPTPerplexityClaudeCopilotGemini

Last checked 2026-06-05

What is Position-Adjusted Word Count?

Position-Adjusted Word Count (PAWC) is the metric introduced in Aggarwal et al. 2023's GEO paper1 to measure how prominently a given source appears in a generative AI engine's answer. It scores the share of the answer's text attributable to a source, but weights each attributed sentence by an exponentially decaying function of its position, so text that appears earlier in the answer counts for more. Despite the name, it is a normalized, position-weighted share, not a raw count of words. The paper also reports a plain Word Count (the same share without the position weighting) and a GPT-judged Subjective Impression; PAWC is the one nearly every GEO claim is built on. On the paper's main GEO-bench an unoptimized page scores PAWC 19.5, and the well-known "GEO can boost visibility by up to 40%" figures are improvements over that baseline.

To see the position weighting concretely: suppose an answer is ten sentences long and three of them are drawn from your page. If those three are the 1st, 2nd, and 3rd sentences, your PAWC is markedly higher than if they are the 8th, 9th, and 10th, even though the plain word-count share is identical. The metric bakes in the assumption that earlier text in an answer is more prominent, an assumption calibrated on 2023 GPT-3.5 responses that may not hold the same way on a 2026 engine.

The distinction that matters: PAWC measures word-count share, not citation rate or ranking. A page can lift its PAWC (a bigger, earlier slice of the answer's words) without being cited more often, or cited at all as a named source. Reading a PAWC gain as a "citation lift" is probably the most common misinterpretation of the GEO literature.

Status in 2026

PAWC is the de facto metric behind the GEO paper's reported content-method effect sizes: nearly every headline number in the GEO marketing literature (Quotation Addition +43%, Keyword Stuffing -8.7%, and the rest) is a PAWC delta on Aggarwal's 2023 GEO-bench. Three caveats ride under the headline numbers and usually get dropped. It is a single-actor synthetic measurement (one page is optimized while the rest of the web holds still); it was run on GPT-3.5-turbo in 2023; and the "30-40%" headline is the main-bench figure for three named methods, while the paper's own Perplexity.ai per-engine table tops out near +22%.

The 2025 C-SEO Bench benchmark2 made the word-count-versus-citation gap explicit. It re-scored seven of Aggarwal's methods by citation ranking (which source the engine cites first) under multi-actor conditions and found the tested methods "do not generally improve citation ranking," noting that Aggarwal's own PAWC data, read for citation ranking, points the same way. So PAWC remains the field's reference metric, but it measures a different thing than the citation outcome most practitioners actually want.

How to apply

When you meet a GEO claim, read the metric, not just the percentage:

  • Translate PAWC into "share of the answer's words," not "citation lift." A "+40% PAWC" page occupies more of the answer's text, positioned earlier; it does not mean the page is cited 40% more, or named as a source at all. Ask whether the claim is about word-share or about being the cited source, because they are different outcomes.
  • Anchor every PAWC figure to the 19.5 baseline and its testbed. The numbers are deltas over an unoptimized page on a 2023 GPT-3.5 benchmark; treat them as directional signal under those conditions, not a multiplier for any 2026 engine. Per-engine results vary (the paper's Perplexity table tops near +22%, not 30-40%).
  • Measure the citation outcomes PAWC cannot see. Whether you are the named, linked source is the outcome PAWC's word-count share leaves out; track it directly with a probe panel, or with practitioner metrics like citation share and cite-ability.

What to skip: quoting a single PAWC number as proof a tactic "works for AI citation." The metric is word-count prominence in a 2023 single-actor test, and the 2025 multi-actor re-test found the same tactics do not move citation ranking.

How it relates to other concepts

  • Generative Engine Optimization is the practice PAWC was built to measure; nearly every GEO effect size in the literature is a PAWC number.
  • GEO content methods are the techniques scored in PAWC; that cluster pillar's evidence table is entirely PAWC deltas.
  • C-SEO Bench is the 2025 counter-benchmark that scores by citation ranking instead of PAWC and reaches opposite conclusions for most methods, the clearest demonstration that PAWC is not citation.
  • Citation share is a practitioner citation metric (how often you are the cited source); it measures the outcome PAWC's word-count share leaves out.
  • Cite-ability is the practitioner content property (extractable, self-contained, front-loaded passages) that may help explain why some pages score higher under PAWC-style measurement.

Footnotes

  1. Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, Deshpande. "GEO: Generative Engine Optimization." arXiv:2311.09735, November 2023 (KDD 2024). Princeton + IIT Delhi + Georgia Tech + Allen Institute for AI. PAWC (the paper's notation is Imp'_wc) is the position-adjusted share of the response attributable to a source: each attributed sentence's word count is multiplied by an exponentially decaying function of its position (e^(-pos/|S|): a larger multiplier for earlier sentences, a smaller one for later ones), then normalized over the whole response. The paper also reports a plain Word Count (Imp_wc, no position term) and a GPT-judged (G-Eval) Subjective Impression (relevance, influence, uniqueness, position, content amount, click likelihood, diversity). No-modification baseline PAWC on the main GEO-bench is about 19.5; Table 1 method values include Quotation Addition 27.8, Statistics Addition 25.9, Keyword Stuffing 17.8. Abstract headline: "GEO can boost visibility by up to 40% in GE responses" (an aggregate; an individual method's Table 1 gain can exceed it, e.g. Quotation Addition +43%). Testbed: GPT-3.5-turbo, top-5 Google sources, 2023. Formula + baseline primary-source verified 2026-06-05 against the ar5iv HTML mirror of arXiv:2311.09735. ↩

  2. Puerto, Gubri, Green, Oh, Yun. "C-SEO Bench: Does Conversational SEO Work?" arXiv:2506.11097, NeurIPS 2025 Datasets & Benchmarks Track. Re-tests Aggarwal's methods scored by citation ranking (not PAWC) under multi-actor conditions; headline "most current C-SEO methods are largely ineffective, contrary to reported results in the literature" (the abstract adds they "frequently have a negative impact on document ranking," its phrasing for the same source-rank outcome this entry calls citation ranking), and the Discussion notes Aggarwal's own PAWC data "implicitly" indicates the methods do not generally improve citation ranking. ↩

Part of Methodology· editorial cluster, not a semantic link

Also in this cluster: Citation probe protocol · External traffic disambiguation

Related terms

FAQ

Does a higher PAWC mean my page gets cited more?
No, and this is the most common misreading of the GEO literature. PAWC measures the position-weighted share of an answer's words drawn from your page, not how often the engine names or links you as a source. A page can occupy a bigger, earlier slice of the answer text (higher PAWC) without being cited more, or cited at all. If you care about being the named source, measure citation directly (a probe panel, or metrics like citation share), not PAWC.
What is the baseline PAWC and what do the GEO percentages mean?
On Aggarwal et al. 2023's main GEO-bench, an unoptimized page scores PAWC about 19.5. The headline numbers (for example 'up to 40%') are relative improvements over that baseline, measured on GPT-3.5-turbo with top-5 Google sources in 2023. They are directional signal under those single-actor synthetic conditions, not a multiplier you can expect on a 2026 engine; the paper's own Perplexity.ai table tops out near +22%, not 30-40%.
How is PAWC different from plain Word Count?
The paper reports both. Plain Word Count is the share of the answer's words attributable to a source. Position-Adjusted Word Count multiplies each attributed sentence by an exponentially decaying function of its position, so text that appears earlier in the answer counts for more, on the reasoning that earlier placement is more prominent. The paper also reports a third, GPT-judged 'Subjective Impression' metric. PAWC is the one nearly every reported GEO effect uses.

Sources & further reading

Get the monthly digest

New terms shipped that week, plus one observation from the AI-citation tracker.

More about what you'll get