The Attention Economy

The Analytics Lie You're Paying For

    You're tracking the wrong metrics. Your entire analytics stack—Google Analytics, Mixpanel, Amplitude—is built on a lie.





    Here's the lie: **"The last thing a user clicked before converting is what caused the conversion."**





    This is called last-click attribution, and it's the foundation of nearly every marketing analytics tool ever built.





    But here's what sequential recommendation models—built by Amazon, Alibaba, and Google—actually show:




    - Attention focuses on 2-3 recent items, not entire history (SASRec, 2018)
    - Time gaps matter more than position (TiSASRec, 2019)
    - ≥75% of "successful recommendations" are substitution, not discovery (Amazon research, 2015)
    - Last-click attribution systematically misattributes value by ignoring the sequence that built intent

The Truth

      The ad you paid $5 per click for probably didn't cause the sale. It just happened to be the last thing they clicked before they bought what they were already going to buy.




      Your analytics say the ad "worked." The science says it was **credited for traffic it intercepted, not created**.

What Sequential Models Actually See

Traditional Analytics (Last-Click)

    **What it tracks:** User saw ad → clicked → bought

    **Conclusion:** Ad caused purchase (give ad full credit)





    **What it ignores:** User's entire behavioral history, time gaps between actions, intent that was building before ad appeared, context of what user was already doing





    **Example:**




    - Day 1: User searches "thriller books like Gone Girl"
    - Day 2: User browses 20 thriller titles, adds 3 to wishlist
    - Day 3: User sees ad for your thriller, clicks ad, buys book




    Last-click attribution verdict: Ad caused 100% of sale ($5 CPC was worth it!)





    But was the ad the cause, or did it just intercept a user who was already in-market for thrillers?

Sequential Models (Attention-Weighted)

    Sequential models track entire sequence of actions with attention weights showing which past actions influence current decision:




    - Day 1 search: 0.35 attention weight
    - Day 2 browsing: 0.42 attention weight
    - Day 3 ad click: 0.15 attention weight




    **Attention-weighted attribution verdict:**




    - Browsing session (Day 2): 42% credit
    - Search intent (Day 1): 35% credit
    - Ad click (Day 3): 15% credit
    - Other factors: 8% credit




    The ad didn't "cause" the purchase—it was the **last step in a sequence** where most intent was already formed. The ad facilitated conversion for someone who was already 77% of the way to buying.





    And you paid $5 to claim credit for traffic that was already coming.

The Research: How Sequential Models Exposed the Lie

Study 1: SASRec (Self-Attentive Sequential Recommendation)

    **Source:** Kang & McAuley, "Self-Attentive Sequential Recommendation" (2018)





    Built a Transformer-based model that learns which past actions influence current decisions. Trained on Amazon product purchases, Steam game plays, MovieLens ratings.





    **Key finding:** "Attention consistently focuses on 2-3 recent items. Items beyond the most recent 5 typically receive less than 5% attention weight."





    When you're deciding whether to buy a book, your brain is mostly looking at: what you browsed in the last session, what you added to your wishlist recently, what genre you've been binging this week.





    Your purchase 6 months ago? Your search 2 weeks ago? **Algorithmically irrelevant.** Attention has decayed to near-zero.

Study 2: TiSASRec (Time-Aware Sequential Recommendation)

    **Source:** Li et al., "Time Interval Aware Self-Attention for Sequential Recommendation" (2019)





    Added explicit modeling of time gaps between actions. Discovered that **temporal distance matters more than positional distance**.





    **Key finding:** "A click 10 minutes ago is more influential than a click that was 3 positions ago but occurred 2 days ago. Time intervals explain 15-25% of variance that position alone misses."





    Last-click attribution doesn't just ignore history—it **doesn't understand which history matters**. A click 2 days ago is not the same as a click 2 hours ago.

Study 3: Causal Impact (Amazon Research)

    **Source:** Sharma, Hofman, Watts, "Estimating the Causal Impact of Recommendation Systems from Observational Data" (2015)





    Used natural experiments (viral tweets, TV mentions, external traffic spikes) as instruments. Compared what users did after recommendations vs. what they would have done without recommendations.





    **Key finding:** "At least 75% of clicks on recommended items would have occurred through other paths (search, direct navigation, external links). Recommendations get credit for traffic they redirected, not created."

The Devastating Implication

      When you pay for an ad, click it, and buy—**your analytics assume the ad created 100% of the value**. But if sequential models show that 77% of your intent was already formed, and causal analysis shows that 75% of clicks are substitution...




      **You're paying for credit, not for causation.**

Real-World Example: Author Ad Campaign

    **User journey:**




    - Day 1, 10:00 AM: Search "best psychological thrillers 2025"
    - Day 1, 10:15 AM: Browse 15 thriller titles, read 8 sample chapters
    - Day 1, 10:45 AM: Add 3 books to wishlist (yours is one)
    - Day 2, 8:00 PM: See Facebook ad for your book
    - Day 2, 8:05 PM: Click ad, go to Amazon
    - Day 2, 8:10 PM: Buy your book




    **Traditional analytics report:**




    - Last-click attribution: Facebook ad caused 100% of sale
    - Cost: $4.50 per click
    - ROAS: Book is $4.99, you earn $3.50, breakeven is 0.78 ROAS
    - Verdict: Profitable campaign! Scale it!




    **Sequential model analysis (what actually happened):**





    Attention weights:




    - Search + browsing session (Day 1): 0.60 attention (primary driver of intent)
    - Wishlist add (Day 1): 0.25 attention (strong interest signal)
    - Ad click (Day 2): 0.10 attention (reminder/convenience, was already in wishlist)
    - Other factors: 0.05 attention




    Causal analysis: User already wishlisted your book (high likelihood of purchase even without ad). Probability user would have bought anyway: ≥70%





    **True causal lift:** 30% or less





    **Revised ROAS:** Cost $4.50, but 70% of that revenue would have happened anyway. Incremental revenue: $3.50 × 0.30 = $1.05. Actual ROAS: $1.05 / $4.50 = **0.23** (you're losing $3.45 per click)





    Verdict: Unprofitable campaign burning money claiming credit for organic traffic.

The Metrics You Should Actually Track

1. Intent Transition Tracking

    Instead of asking "What did they click last?" ask "What sequence built their intent?"





    **Example dashboard:**




    - 45% of buyers: Search → Browse → Purchase (single session, high intent from start)
    - 30% of buyers: Browse → Wishlist → Ad Click → Purchase (intent built over 2-3 days)
    - 15% of buyers: Recommendation → Sample → Purchase (discovery-driven)
    - 10% of buyers: Direct link → Purchase (external traffic)




    **Actionable insight:** 45% of your buyers have high intent from their first search. You don't need ads to convert them—you need good SEO and metadata. Focus ad spend on the 30% who need reminders.

2. Time-Weighted Engagement

    Not just "did they engage?" but "how quickly did they engage?"





    **Engagement velocity:**




    - Fast path (same session): Search → Browse → Purchase within 30 min = High intent, quality match
    - Slow path: Browse → Wishlist → 7 days gap → Ad → Purchase = Needed reminder or price drop
    - Very slow path: Recommendation → 30 days → Purchase = Likely substitution

3. Substitution vs. Discovery Ratio

    For each purchase, calculate intent pre-score:




    - Searched for genre: +0.35 intent
    - Browsed 10+ titles: +0.30 intent
    - Wishlisted book: +0.25 intent
    - Pre-ad intent score: 0.90 (already 90% likely to buy)




    Then user sees ad → buys




    - Substitution score: 0.90 (ad claimed credit for 90% pre-existing intent)
    - Discovery score: 0.10 (ad added 10% incremental value)




    **Aggregate across campaigns:** Campaign A with 0.75 substitution is mostly stealing credit. Campaign B with 0.30 substitution is mostly creating demand. Campaign B is 3x more valuable.

4. Attention-Weighted Attribution

    Distribute conversion value across touchpoints based on attention weights:





    Traditional (last-click): Facebook ad = 100% credit = $3.50 revenue





    Attention-weighted:




    - Search (Day 1, attention 0.35): 35% credit = $1.23
    - Browse session (Day 1, attention 0.30): 30% credit = $1.05
    - Wishlist (Day 1, attention 0.25): 25% credit = $0.88
    - Facebook ad (Day 2, attention 0.10): 10% credit = $0.35




    **Actionable insight:** Your SEO (which drove the search) created $1.23 of value. Your metadata (which drove browsing) created $1.05. Your ad created $0.35. Invest accordingly.

How Teneo Measures What Actually Matters

    Most platforms use last-click because it's easy and makes their ads look better. We use sequential models because they're accurate.

1. Full Sequence Tracking

    We track every reader interaction from discovery through purchase with time gaps, engagement depth, and cross-session patterns. We know whether your sale came from a coherent high-intent session or a slow 30-day substitution path.

2. Attention-Based Analytics

    We show you which past actions influenced each purchase with attention weights, time decay visualization, and substitution vs. discovery breakdown.





    **Example dashboard:**




    - 40% fast-path (search → buy within 1 session) → Optimize metadata
    - 35% slow-path (browse → wishlist → reminder → buy) → Optimize email nurture
    - 15% discovery (recommendation → sample → buy) → Optimize for read-through
    - 10% external (social link → buy) → Optimize author platform

3. Intent Transition Funnels

    We track state transitions (Unaware → Aware → Interested → Intent → Purchase) with time in each state and which interventions move people between states.

4. Causal Lift Estimation

    For each marketing intervention, we estimate counterfactual (what would have happened without it) using natural experiments and cohort matching. We report true incremental lift, not substitution.





    Traditional analytics: Email campaign drove $175 revenue

    Causal lift analysis: Treatment group 50 purchases, control group 38 purchases. Incremental lift: 12 purchases. True incremental revenue: $42 (not $175)

What This Means for Authors

Three Takeaways

      **1. Most of Your "Successful" Marketing Is Claiming Credit for Organic Traffic**

      If 75% substitution is real, you're probably wasting money on ads that intercept readers who were already coming. Audit your sales: How many came from high-intent search vs. low-intent discovery? If greater than 60% are high-intent, focus on SEO/metadata, not ads.





      **2. Time Gaps Reveal True Interest**

      Readers who buy within 24 hours of discovery are your best fit. Readers who take 30 days needed reminders, not discovery. Segment by time-to-purchase and optimize differently for fast vs. slow buyers.





      **3. Platforms That Use Last-Click Are Lying to You**

      If their analytics show "last-click attribution," they're systematically over-crediting interventions. This makes their ads/promotions look better than they are. Demand attention-weighted or multi-touch attribution.

The Attention Economy

The Analytics Lie You're Paying For

The Truth

What Sequential Models Actually See

Traditional Analytics (Last-Click)

Sequential Models (Attention-Weighted)

The Research: How Sequential Models Exposed the Lie

Study 1: SASRec (Self-Attentive Sequential Recommendation)

Study 2: TiSASRec (Time-Aware Sequential Recommendation)

Study 3: Causal Impact (Amazon Research)

The Devastating Implication

Real-World Example: Author Ad Campaign

The Metrics You Should Actually Track

1. Intent Transition Tracking

2. Time-Weighted Engagement

3. Substitution vs. Discovery Ratio

4. Attention-Weighted Attribution

How Teneo Measures What Actually Matters

1. Full Sequence Tracking

2. Attention-Based Analytics

3. Intent Transition Funnels

4. Causal Lift Estimation

What This Means for Authors

Three Takeaways

Further Reading

Try Analytics That Actually Measure Causation