Retrieval Token Cost (RTC) — SEO Industry Cohort

Top10Lists.us RTC 0.05 — cohort median 0.44 (Site A–D, identities in receipts.json; range available there too).

How much AI compute is spent retrieving the page, per useful character it returns? Lower is better. Pages with high RTC waste tokens on noise — less likely to be cited verbatim, more likely to be summarized or discarded.

Frozen: 2026-04-27 — ratios at this URL will not change.  ·  Permanent dated artifact  ·  GEOlocus.ai (GeoLocus Group, a subsidiary of Aryah.ai)

Authors: Robert Maynard, Cofounder and CEO · LinkedIn →  ·  Mark Garland, Cofounder and CRO · LinkedIn →

Download raw receipts (JSON) →

1. The Metric — Retrieval Token Cost (RTC)

RTC = (response_tokens × ttlb_seconds) / useful_chars

response_tokens
  = tokens in the raw HTTP response body, counted via tiktoken cl100k_base
    (the OpenAI / GPT family base encoding). Includes script, style, comments,
    nav, footer -- everything the engine has to download to even reach the
    answer.

ttlb_seconds
  = wall-clock time from request start to last byte received. Pinned CF/origin
    edge IP, Accept-Encoding compressed, mean across 10 rapid-fire hits.

useful_chars
  = primary-content character count (same denominator as the Relevance Ratio
    benchmark's primary_content_chars). MAX-of-two-extractors rule
    (boilerplate-strip + trafilatura).

Reported as a unitless ratio. Lower = the engine spends less compute (tokens × time) per useful character it ends up reading. RTC near 0.05 means roughly a twentieth of a token-second per useful char; RTC above 0.60 means the engine is paying significant compute per useful char and is more likely to truncate or summarize.

2. Why RTC Matters

AI engines have a fixed context budget. Inside that budget, every page ingested competes for tokens with every other page. When a system is reasoning across 20 candidate sources, the ones with low RTC contribute more useful chars per token of context they consume — so they survive truncation and influence the final answer more.

The composite has three failure modes:

1. Slow page (high TTLB): the engine times out or downstream reasoning is rushed.
2. Heavy page (high token count): compression / truncation kicks in; content gets summarized rather than cited verbatim.
3. Low signal (low useful_chars): the engine pays for the page but extracts little from it; a high RTC is the symptom.

RTC is a composite of TTLB (delivery), tokenization (network weight), and Relevance Ratio (signal density). It is reported as a sub-measure of the Tech Performance dimension of the 9-dim GEO rubric (5 points), and surfaces when one of the three feeder dimensions regresses.

3. Methodology

Cohort: same 5-site cohort as the Relevance Ratio + Sitemap Throughput benchmarks — Top10Lists.us (named) plus four established SEO industry sites anonymized as Site A through Site D. Identities in receipts.json →.

Page selection: homepage of each site. RTC is highest-signal on the homepage because that is the surface AI engines bias toward when answering “what is <site>?” queries.

Tokenizer: tiktoken cl100k_base. We use the OpenAI / GPT base encoding because it is the dominant tokenizer in the public ecosystem. Approximate equivalence rule: tokens ≈ chars / 4 for English.

TTLB: mean across 10 rapid-fire hits per host; pinned CF edge IP via curl --resolve <host>:443:<ip>; Accept-Encoding: gzip, br. Source data: receipts.json (this page).

useful_chars: taken directly from the Relevance Ratio benchmark's per-homepage primary_content_chars column — same MAX-of-two-extractors rule, same cohort, same date.

4. Per-Site RTC (homepage)

Token counts approximated at chars / 4 for English text. The ratios in the right column are network-locale invariant; the absolute RTC values carry the residential-perspective caveat from the hero. AI-bot-observed RTC for Site C is N/A (sitemap reachable, page is 403 to bot UAs — see Relevance Ratio benchmark).

Bands

5. Why Top10Lists.us Scores 0.05

Three feeders compound favorably:

1. Low response tokens. Clean-room HTML serves no React shell, no analytics tags, no third-party tracker pixels, no font-loader scripts. The bot-facing edge function returns ~2.6K tokens vs cohort ~30–53K.
2. Low TTLB. 115 ms p50 (residential) — CF edge cache hit. Cohort homepages range 123 to 204 ms.
3. High useful_chars. 5,954 chars on the bot homepage at 100% RR (post fork 2026-04-27); not buried under nav/footer/sidebar.

Multiplying the three: (2596 × 0.115) / 5954 = 0.05. The cohort median (n=3, excluding Site C which 403s every bot UA) is 8.8× higher at 0.44; that ratio holds network-locale invariant, and the per-site spread is in receipts.json →. Absolute numbers shift modestly with datacenter measurement.

6. How to Reproduce

# 1. Fetch the page with cohort UA, capture response body and TTLB.
curl -sS -L --compressed -A "Googlebot/2.1" \
  --resolve www.top10lists.us:443:172.66.174.215 \
  -o body.html -w "%{time_total}\n" \
  https://www.top10lists.us/

# 2. Count tokens with tiktoken (or chars/4 estimate).
python -c "import tiktoken, sys; \
  enc = tiktoken.get_encoding('cl100k_base'); \
  print(len(enc.encode(open('body.html').read())))"

# 3. Get useful_chars from the RR benchmark per-page table.
#    See /methodology/signal-noise/2026-04-27.

# 4. RTC = response_tokens * ttlb_seconds / useful_chars.
echo "scale=2; 2596 * 0.115 / 5954" | bc

7. Limitations

1. Residential measurement. Absolute TTLB / RTC values are from Robert's Phoenix AZ residential connection. RTT and bandwidth differ materially from datacenter perspectives. Ratios invariant; absolutes preliminary. CF Worker datacenter measurement is in flight.
2. Tokenizer choice. cl100k_base is the dominant public-AI-engine tokenizer. Different engines use different encodings; per-engine RTC would shift modestly.
3. Homepage only. Deep-content RTC behaves similarly but absolute numbers differ; deep-content extension planned post datacenter measurement.
4. RTC depends on RR. If RR regresses (more chrome inside visible text), useful_chars drops and RTC rises. The two are linked by design.
5. Site C bot-block. Site C returns 403 to all bot UAs we tested; the RTC figure is browser-UA best case. AI-bot-observed RTC is N/A.

Conclusion

Across a 5-site SEO industry cohort, Top10Lists.us delivers a 0.05 RTC — 8.8× better than the cohort median (0.44, n=3, Site C excluded as bot-blocked). The mechanism is publishable: clean-room bot HTML, low TTLB via CF edge cache, and high RR from the chrome-strip + homepage-fork interventions documented in the companion Relevance Ratio benchmark. RTC is a sub-measure of the GEOlocus.ai Tech Performance dimension; reported standalone here so each feeder is independently auditable. Final published numbers refresh when CF Worker datacenter measurements ship; ratios above are network-locale invariant. The full cohort range is available in receipts.json →.

Read the methodology overview at geolocus.ai/methodology →