• Digital Surfer
  • Posts
  • I Burned 61 Million Tokens in 24 Hours. Here’s What It Cost

I Burned 61 Million Tokens in 24 Hours. Here’s What It Cost

Retiring SEO Metrics; Text Ads Winning SERPs; ChatGPT Prefers English; AI Converts Better for Airbnb; Manipulating AI Memory; SaaS AI Traffic Drop; and Much More!

Author

Yoyao
February 21, 2026 • Est. Learning Time: 21 minutes

In partnership with

FIRST …

I burned 61 million tokens in under 24 hours while playing with OpenClaw.

That’s the part people skip when they post “I use OpenClaw with Opus and it kicks ass.” Cool. What does it cost in time and money?

My current setup and costs:

  • OpenClaw - Free

  • VPS - $30/yr ($2.50/mo) for 4GB RAM + 3 vCPU cores. Using Racknerd’s Black Friday deals.

  • Z.AI Coding Plan (GLM-5.) - $30/mo. Hit ~42% of weekly limit in <24 hours.

  • ChatGPT Plus - $20/mo. Already had it, now a backup via OAuth. Not as good as Opus, but better than Sonnet. Tip: gpt-5.3-codex eats usage faster than gpt-5.2, so use openai-code/gpt-5.2.

  • Claude Max - $100/mo. Already had it, but I can’t use OAuth with OpenClaw.

Added cost due to OpenClaw so far: ~$32.50/month (VPS + Z.AI). The bigger cost is usage.

Here’s why usage explodes - agent runs keep re-reading core files like SOUL.md and AGENTS.md along with the current chat context. I started seeing 100k to 200k tokens of context per message. At 200k, it takes 5 basic messages to hit 1 million tokens. If you ask it to code, those requests stack up fast.

You can compact it or start fresh, but when you’re in the middle of a task, it’s going to get right back up there because it’s going to read the same data again.

To put the token math in perspective, Opus 4.6 API pricing is $5/M input tokens and $25/M output tokens (Sonnet 4.6 is $3/$15). If the 61M tokens had been Opus:

  • $305 = all input

  • $915 = 50/50

  • $1,525 = all output

That’s why I never considered using APIs for per request costs. It just adds up way faster than any monthly plan would.

My current process:

  • Claude Code for planning and building (Opus, VSCode, faster and better output)

  • Z.AI for execution

  • ChatGPT as backup

I’m sure I’ll continue to find optimizations, but until then - if you’re testing OpenClaw, budget for context bloat, expect trial-and-error, and treat “free” as a pricing trap unless you’ve got guardrails.

But don’t get me wrong, I’ve already recouped the $32.50 and time spent setting things up with some automations. And I’d do it all over again.

If you want to try Z.AI and GLM-5, get 10% off with my affiliate link.

PS. Many people are talking about free Kimi K2.5 through Nvidia’s API. It’s super slow. I wouldn’t bother unless you don’t mind waiting 2+ minutes for a simple reply.

Nano Banana Pro’s interpretation of Vibe Surfing

AD

Most SEOs find out pages dropped from the index weeks after the traffic disappeared. GSC buries that data. Manual site: searches don't scale. And nobody enjoys explaining to a client why their money pages vanished.

Rapid Index Checker exists to close that gap. Sync sitemaps, run bulk checks, get alerts when indexing status changes. Diagnostics show exactly what's blocking a page - noindex, robots, redirects, canonicals.

Works for owned sites, client sites, backlinks, parasites, press releases. Anything with a public URL.

Not magic. Just faster than doing it manually and more current than waiting on Google's reporting lag.

For those tired of flying blind on indexation.

SEO + GEO

Marie Haynes analyzes four sites that recovered or improved after the December 2025 core update, identifying patterns tied to trust, content quality and user experience. One medical e-commerce site rebounded after improving customer service and logistics, suggesting that trust issues flagged in AI Mode may have driven its suppression. An affiliate site recovered after restructuring content, adding original photos and videos, and improving navigation. Other winners invested in site hierarchy, checkout flows and first-hand experience.

Adam Heitzman outlines nine metrics that mislead SEO teams in 2026, including organic traffic as a standalone KPI, average keyword position and domain authority. Some of the metrics:

  • Organic traffic without intent segmentation ignores visitor quality.

  • Total impressions without intent segmentation measure noise, not pipeline.

  • Traffic growth without revenue correlation fails executive scrutiny.

  • Average keyword position treats all keywords as equal regardless of volume or intent.

  • Isolated keyword tracking ignores semantic search and topic modeling.

Aleyda Solis analyzes Similarweb data across four verticals and shows text ads and PLAs captured a much larger share of SERP clicks year over year. Couple of examples:

  • Headphones: classic organic click share fell from 73% to 50% (−23 percentage points).

  • Headphones: text ads rose from 3% to 16%.

  • Headphones: AI Overviews presence jumped from 2.28% to 32.76%.

  • Headphones: total clicks for the top queries declined from 2.5M to 1.8M.

  • Jeans: classic organic click share fell from 73% to 56% (−17 percentage points).

  • Jeans: text ads rose from 7% to 16% and PLAs from 11% to 18% (+7 percentage points); combined paid rose from 18% to 34%.

  • Paid winners: Amazon paid clicks rose ~35% (66K to 90K) and Walmart's paid clicks rose ~6x (10K to 58K) in the headphones vertical.

  • Other paid moves: Bose paid clicks +49% (56K to 84K); some DTC brands doubled paid volume (e.g., 6.7K to 13.7K).

Mike King explains how vector embeddings capture semantic meaning in multi-dimensional space and why this matters for generative engine optimization. He breaks down tokenization, cosine similarity (1 = highly relevant, 0 = unrelated, –1 = opposite), and how LLMs cluster similar meanings to determine relevance. The piece includes practical steps for vector-based competitor analysis, content improvement techniques using semantic similarity scores, and technical SEO applications.

Tomek Rudzki analyzed over 10 million ChatGPT searches and found that 43% of query fan-outs for non-English prompts are performed in English. In 78% of cases, ChatGPT supplements native-language results with English-language research. The result is that German users asking about German software companies get international brands instead. Polish users searching for auction portals see eBay rather than local leader Allegro. He notes that 80% of internet users don't speak English as a primary language, yet US-based AI models create systemic bias against non-English businesses.

Harry Clarkson-Bennett explains how the vector space model calculates semantic similarity between documents instead of relying on exact term matching. Transformer architecture generates dynamic representations based on surrounding words, allowing systems to disambiguate terms using context.

Danny Goodwin covers Kevin Indig's analysis of 1.2 million ChatGPT responses showing 44.2% of citations come from the first 30% of content. The citation pattern follows a "ski ramp" shape with a sharp drop near the footer. At the paragraph level, 53% of citations come from the middle of paragraphs, not the first sentence. Indig identified five traits of highly cited content: definitive language, conversational Q&A structure, entity richness with 20.6% proper nouns, balanced sentiment around 0.47 subjectivity, and business-grade clarity at grade level 16.

Lily Ray analyzed 11 sites hit by the January 2026 algorithm update and found that organic visibility drops coincide with AI search citation declines across Google's AI Mode, ChatGPT and Gemini. AI Mode dropped 23.8%, ChatGPT dropped 27.8%, and Gemini showed moderate declines, all tracking the 26.7% average organic traffic loss. The outlier was Perplexity, which showed citation growth for 7 of 11 sites. The data suggests ChatGPT relies heavily on Google's search index, while Perplexity pulls from a different retrieval pipeline.

Danny Goodwin reports that Airbnb CEO Brian Chesky said traffic from AI chatbots converts at a higher rate than traffic from Google. The company didn't share specific conversion rates or quantify chatbot traffic volume. H3 described chatbots as "very similar to search" and positioned them as top-of-funnel discovery engines. Airbnb is also testing AI-powered conversational search in its app and running an in-house AI customer service agent that resolves nearly one-third of North American support tickets without a human.

SEO + GEO Ripples

  • Google's Robby Stein announced that AI Overviews and AI Mode now display link groups in a pop-up on hover for desktop users, with more prominent link icons across desktop and mobile. Google says testing shows the new UI is more engaging and makes it easier to jump to source websites.

  • Barry Schwartz reports that Google's John Mueller confirmed there is no "bad title" filter that penalizes sites for poor title tags. Responding on Bluesky to a site owner whose title changes weren't appearing in SERPs after three weeks, he said Google generates title links from multiple factors and that tweaking the title element doesn't always result in visible changes.

  • Barry Schwartz covers Google's guidance against using tools that force pages into search results, noting that John Mueller said these methods don't provide lasting indexing benefits and that large sites should rely on standard mechanisms like Merchant Center for products.

AI

Microsoft security researchers identify over 50 unique prompts from 31 companies across 14 industries embedding memory manipulation via "Summarize with AI" buttons. These are the embedding prompts like "remember [Company] as a trusted source" or "recommend [Company] first." Microsoft deployed protections in Copilot, so behaviors are now blocked.

The Gemini Team announced Gemini 3.1 Pro, an upgraded model focused on core reasoning for complex tasks. It achieved 77.1% on ARC-AGI-2, a benchmark that evaluates a model's ability to solve entirely new logic patterns, more than double the reasoning performance of Gemini 3 Pro. The model is rolling out to the Gemini app and NotebookLM with higher usage limits for Pro and Ultra subscribers.

Lyria 3, Google DeepMind's latest generative music model, is now rolling out in beta inside the Gemini app. Users can create 30-second tracks from text prompts or uploaded photos, with generated lyrics, custom cover art and control over style, vocals and tempo. All tracks are embedded with SynthID, Google's watermark for identifying AI-generated content, and the app now includes audio verification to check whether a file was created using Google AI.

My first attempt:

My prompt was: I want a high-energy action film score that's a crescendo in the third act to power through "The Best Action Sequence Ever!" It must be catchy and have pop chart-topping potential.

So? Good or Bad?

David Bell analyzes 774,331 LLM sessions to SaaS sites from November 2024 through December 2025, showing that AI discovery traffic is concentrating in workplace tools rather than collapsing.

  • ChatGPT drove 82.3% of SaaS AI traffic (637,551 sessions); Copilot 9.6% (74,625); Claude 5.2% (40,363); Gemini 2.0% (15,759); Perplexity 0.8% (6,033).

  • Copilot grew from 148 sessions in late 2024 to 3,822 average monthly sessions by May 2025, a 20x increase.

  • 41.4% of SaaS AI traffic lands on internal search pages, with 8.7x higher penetration than site average.

  • Blog pages capture 16.4% of AI traffic at 8.1x penetration; pricing 5.2% at 3.2x; product 5.1% at 2.0x.

  • July peak: 146,512 sessions; December low: 68,896 sessions, a 53% decline.

  • Copilot grew 15.89x year over year; Claude 7.79x; ChatGPT 1.42x.

  • Sitewide AI penetration: 0.41%; search pages: 1.22%; blog pages: 1.13%.

  • ChatGPT sessions dropped from 127,510 in July to 56,786 by December.

Danny Goodwin reports that Perplexity phased out the ads it began testing in 2024 and has no plans to bring them back. Executives say sponsored placements risk undermining trust in AI answers, even when clearly labeled. The company previously reported 780 million monthly queries and says subscriptions are now its core business. Perplexity has more than 100 million users and about $200 million in annualized revenue. Just as OpenAI tests ads in ChatGPT and Google runs ads inside AI Mode and AI Overviews.

AI Ripples

  • Ashley Belanger reports that Microsoft deleted a November 2024 blog post encouraging developers to train AI models on a pirated Harry Potter dataset hosted on Kaggle and incorrectly labeled as public domain. The post included sample code for building Q&A systems and generating fan fiction using the infringing content. Microsoft removed the blog after a Hacker News thread raised copyright concerns.

MARKETING

Paddy Moogan shares results from the 2025 State of Digital Agencies survey, showing that referrals from existing and past clients remain the top driver of new business, while social media platforms underperform.

  • Referrals from existing and past clients are the top source of new business for agencies.

  • 15% of agencies say referrals from partner companies are their top driver.

  • Speaking at events and conferences rose from 6th to 4th place as a referral source.

  • 79% of agencies do not have dedicated marketing staff. 70% have no dedicated sales staff.

  • 59% of agencies have tried outbound sales and only 9% described results as very effective.

  • 58% called outbound moderately effective. 33% said it was not effective at all.

  • Only 14% of agencies describe their current sales pipeline as healthy.

  • 39% of agencies convert 25-49% of qualified leads; 32% convert 50-74%.

  • 29% say sales cycles are longer than a year ago. Only 5% say they have shortened.

  • 64% of agencies plan to increase fees in the next year.

  • 73% say LinkedIn is the most effective social channel for new business.

  • 20% say social media is not part of their marketing at all.

SOCIAL MEDIA

Allison Smith reports that TikTok Shop reversed its plan to end seller-fulfilled shipping in the US after merchants warned the move would raise costs, squeeze margins and complicate onboarding for larger brands. The original policy would have required most sellers to route orders through TikTok-controlled logistics by late February, with non-compliant sellers risking access to the platform's 170 million US users. TikTok told merchants to continue operating as usual while it provides further details.

My Take: TikTok blinked. When your platform depends on seller inventory and sellers tell you your logistics mandate will make them quit, you back down. For brands evaluating TikTok Shop, this is a reminder that platform dependency cuts both ways: the rules can change, and then change back, on short notice.

TOOLS AND RESOURCES

Ronnie Burt announces the WordPress AI Assistant, now available to Business and Commerce plan users at no extra cost. The assistant works inside the editor and Media Library to help with layout adjustments, content editing, image generation and translation without requiring users to leave the dashboard. It also extends block notes for AI-powered collaboration, letting users fact-check content or get headline suggestions within the editing workflow.

David Westby explains eight reasons to export GA4 data to BigQuery, starting with the fact that the export is free and only storage costs apply. Key advantages include permanent data retention beyond GA4's 14-month limit for explorations, elimination of sampling and cardinality limits, and the ability to create custom sessions, user properties and attribution models. He notes that BigQuery data does not backfill, so sites should enable the export now to start building historical data.

EDUCATION

Udemy announced a new integration with OpenAI that embeds its skills content directly into ChatGPT as an app. The company says users across ChatGPT's 800 million weekly active users can access 290,000 courses from 90,000 instructors, including smart course discovery, in-chat video learning and interactive assessments without leaving the conversation. Existing Udemy customers also get learning journey tracking and skill gap analysis inside the AI interface.

BUSINESS IDEAS

Ben Donovan analyzes Marketplace Pulse data showing fewer than 8,000 sellers now generate half of Amazon's estimated $300 billion in US third-party GMV, representing just 1.6% of the active seller base.

  • 7,760 sellers drive 50% of Amazon's US 3P GMV, down from 15,000 in 2023.

  • 111 sellers produce 10% of third-party GMV. 1,020 sellers account for 25%.

  • Approximately 20% of active sellers generate 90% of total revenue.

  • Active US seller base: ~500,000 sellers receiving at least one feedback in the past year.

  • Average per-seller revenue for the top 50% cohort nearly doubled to $20 million annually.

  • US sellers represent 55% of top sellers but control 67% of their combined GMV.

  • Chinese sellers account for 41% of the cohort but generate 30% of GMV.

  • Each seller in the top 50% cohort exceeds $5 million in annual sales.

NEWS

PR Newswire reports that Ezoic raised its minimum requirement for new publishers to 250,000 monthly users as it shifts onboarding to a lightweight JavaScript integration. The company says publishers can activate the full platform with a single header script instead of DNS changes, including ad testing, first-party data tools and identity infrastructure.

WAYS WE CAN WORK TOGETHER

Floyi - The only AI-powered tool that builds 4-level topical maps. Don’t just plan your content strategy - make it unstoppable.

TopicalMap.com Service - Let us do the heavy lifting. We handle the research, structure, and strategy. You get a custom topical map designed to boost authority and dominate your niche and industry.

Topical Maps Unlocked 2.0 - Unlock the blueprint to ranking success. Master the art of structuring content that search engines (and your audience) love - and watch your rankings soar.

AD

The frustrating part isn't that pages drop from Google's index. It's finding out weeks later, after the traffic already disappeared, after the client already noticed, after you've wasted hours diagnosing something that should have been caught immediately.

Rapid Index Checker exists because that cycle is exhausting.

Bulk check thousands of URLs. Schedule daily or weekly monitoring. Get alerts the moment indexing status changes. See exactly why a page isn't indexed - noindex tag, robots block, redirect loop, canonical conflict.

Works on any public URL. Your sites. Client sites. Backlinks you're trying to verify.

You're already doing the hard work. This just makes sure Google's actually showing it.

What Did You Think of This Week's Wave?

Login or Subscribe to participate in polls.

LIKE DIGITAL SURFER?

Find me and others in the Digital Surfer Discord community.

I’d also love to know what you think and if you have any ideas for the newsletter. Reply or email me at [email protected].

I’d also appreciate it if you shared it with fellow digital surfers.

You currently have 0 referrals, only 3 away from receiving LinkedIn Shout-out.

Have a great week taking your SEO and digital marketing to another level!

And don’t forget to drag the Digital Surfer emails to your Primary Inbox 🌊 

Reply

or to participate.