Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Can linking metrics alone predict Google rankings?

4,159 views

Published on

A machine-learning-based ranking factor study run on Google Turkey's search results, presented at DigitalZone 2018 by Kevin Richard.
More info on https://zeo.org/digitalzone/en/

Published in: Internet

Can linking metrics alone predict Google rankings?

  1. 1. MERHABA
  2. 2. THE TEAM KEvin Richard CEO @ SEObserver, seo tool suite Sylvain Peyronnet A.I.director @ Qwant,search engine YiğitKonur Baba
  3. 3. NOT YET ANOTHER RANKINGFACTORSTUDY
  4. 4. CAN LINKING METRICS ALONE PREDICT RANKINGS?
  5. 5. GOOGLE.COM.TR + 100,000 NON-BRANDEDKEYWORDS x 100positions + LINKINGMETRICSONLY + MORE THANREGRESSIONTEST
  6. 6. Not nerdy A bit nerdy Very nerdy Nerdiness scale METHODOLOGY RESULTS CONCRETE ACTIONABLE TAKEAWAYS DEFINITIONS TODAY’S PLAN
  7. 7. Ortasehpa Now, let’s grab the linking metrics for each ranking 1 2 3 4 5 This is an “orta sehpa”,coffee table
  8. 8. Linking metrics FROMMAJESTIC Mydomain.com.tr ................... TrustFlow = quality backlinks strength index CitationFlow = all backlinks strength index TF/CF = quality proportion = spam index Referring domains = well… referring domains TrustFlowReferring domains Example:
  9. 9. TRUSTFLOWEXAMPLES • twitter.com : 100 • google.com : 99 • google.com.tr : 51 • sahibinden.com : 53 • n11.com : 30 • hurriyet.com.tr : 72 • oyunkolu.com : 10
  10. 10. Let’s compareranks with ranks • Each SERP is unique : looking at absolute linking metrics values is ludicrous • List the top 100 URLs, grab their metrics • Get a “rank” for each metric : relative values • Create a model that answers the question: “does this url belongs to top1, top3,top5, top10?”
  11. 11. Ortasehpa Now, let’s grab the linking metrics for each ranking 1 2 3 4 5 This is an “orta sehpa”,coffee table
  12. 12. Ortasehpa
  13. 13. Ortasehpa Let’s order this data #1 #2 #3 #4 #6 #6 #8 #5 #7 #10 We obtained our “DomainTF Rank”
  14. 14. SUMMARY Keyword“orta sehpa” : 100 results URLs1 Record relative linking metrics (ranks) for each position2 Repeat 100.000 times w/ different keywords3 You get your ~10M records dataset4
  15. 15. 100K keywords 2M random non-branded records 2M random non-branded records 2M random non-branded records 80% training set 20% test 80% training set 20% test 80% training set 20% test METHODOLOGY Rank Keyword URL URL TF Rank URL RD Rank 1 falan http://url1.com 2 1 2 falan http://url2.com 1 3 3 falan http://url3.com 10 4
  16. 16. TRAINING DATA testing DATA
  17. 17. Keywords sample • Non-branded queries • Randomly chosen içerde, iphone 7, filmler, a spor, 2 kişilik oyunlar, atatürk, akbank direkt, araba oyunu, game of thrones izle, türkiye haritası, xhamster, canlı maç izle, bim aktüel, netspor, uçak bileti, hd porno, a101, dolar, cuma mesajları, ttnet, kısmetse olur, mp3 indir, araba, sex hikayeleri, periscope, sikiş, haber, taraftarium24, mp3, yerli film izle, fa, minecraft, film, haberler, galatasaray, ingilizce türkçe çeviri, ateş ve su, omegle, ntv, porn, süper loto, g, e okul, türk porno, brazzers, güzel sözler, hava durumu, tjk, kırgın çiçekler, survivor, dizi izle, çeviri, oyun, diriliş ertuğrul son bölüm, rüya tabirleri, hız testi, kiralık aşk, poyraz karayel, film izle, canlı skor, game of thrones, film indir, oyunlar, fenerbahçe, ptt kargo, rokettube, porno, araba oyunları, trt 1, sex, son dakika, trt, vatan, iddaa, yutup, faceboook, altın fiyatları, canlı tv izle, tv izle, beşiktaş, hotmail.com, diriliş ertuğrul, son dakika haberleri, kredi hesaplama, bets10, outlook giriş, hotmail, yabancı dizi izle, youtube video indir, xnxx, son dakika haberler, harita, on numara, eşkiya dünyaya hükümdar olmaz, ucuz uçak bileti, olimpiyatlar, gis, euro, kara sevda, tff, müzik indir, ceviri, dolar kuru, turk porno, türkiye, mhrs, taraftarium, torrent oyun, animasyon, paypal, aksis, dağ 2 izle, ccleaner, içerde son bölüm, lg g4, zamunda, o ses türkiye, magazin, teen wolf, zaman, mtv, en iyi filmler, ingilizce çeviri, barbie, son dakika haber, kurtlar vadisi pusu, bir garip aşk, lig tv özet, fuat avni, zenmate, porno film, ayetel kürsi, kurtlar vadisi, türk bayrağı, video indir, youtube video indirme, a, mercedes, the walking dead, türbanlı porno, lig tv maç özetleri, süper loto sonuçları, çevir, sikis, müzik dinle, gta 5, puan durumu, 29 ekim cumhuriyet bayramı, e, outlook aç, redtube, altin fiyatlari, s, aşk sözleri, gogle, dünya haritası, aöl, hadise, tempobet, burçlar, yabancı dizi, canli skor, clash of clans, milli piyango sorgula, gazete, hyundai, porno indir, aleyna tilki, beeg, en uygun, ahmet kaya, otobüs bileti, istanbul, aşk laftan anlamaz, www, komedi filmleri, iphone 6s, istiklal marşı, korku filmleri, kickass, film izle 2016, açık lise, randevu, deprem, recep ivedik 5, konulu porno, seks, mustafa ceceli, bluestacks, bimer, arka sokaklar, düğün dernek 2, oyun indir, iphone 6, ttnet hız testi, lig tv izle, euro 2016, deadpool, torrent, atlasjet, fesbuk, windows 10, samsung j7, sayısal loto sonuçları, türkçe porno
  18. 18. Canliskor 1 2
  19. 19. SOWE DID ACORRELATIONTEST..
  20. 20. THISIS ACORRGRAM
  21. 21. What canwe noticehere?
  22. 22. What canwe noticehere?
  23. 23. Googleis complex asimple regression/Correlationtest is not enough
  24. 24. DECISION TREES ARENEEDED
  25. 25. LET’S CREATEAN”ALGORITHM” Algorithm noun Word used by programmers when they do not want to explain what they did.
  26. 26. LET’S CREATEAN”ALGORITHM” According to its metrics, is this url in the top1, top3, top5, top10 on this keyword? Answers “Yes/No”
  27. 27. LET’S CREATEAN”ALGORITHM” We will put hundreds of decision trees inside, and they will vote for the answer.
  28. 28. How tobe right 99%of the time? 1. Is http://blabla1.com… #1 ? 2. Is http://blabla2.com… #1 ? 3. Is http://blabla3.com… #1 ? 4. Is http://blabla4.com… #1 ? … 99. Is http://blabla99.com… #1 ? 100. Is http://blabla100.com… #1 ? No No No No … No No No No Yes No No … No Correct! Correct! Wrong! Correct! Correct! … Correct! Guess? Expected Result Ask 100 times the same question, always answer“no”, you’ll be wrong only 1% of the time
  29. 29. How tobe right 99%of the time? 1. Is http://blabla1.com… #1 ? 2. Is http://blabla2.com… #1 ? 3. Is http://blabla3.com… #1 ? 4. Is http://blabla4.com… #1 ? … 99. Is http://blabla99.com… #1 ? 100. Is http://blabla100.com… #1 ? No No No No … No No No No Yes No No … No Correct! Correct! Wrong! Correct! Correct! … Correct! Guess? Expected Result Always say “no”, you’ll be wrong only 1% of the time Saçmalama lütfen
  30. 30. How canweeducate ourrobotto avoid this kind of lazy behavior?
  31. 31. 1. DOWNSAMPLING
  32. 32. 2. Raise the penaltyfor each mistake
  33. 33. Teaching the right things https://medium.com/greyatom/what-is-underfitting-and-overfitting-in-machine-learning-and-how-to-deal-with-it-6803a989c76
  34. 34. Random? How well do weperformvs totalrandomness? Can a dice be moreaccurate thanour MODEL?
  35. 35. Underthe hood Input data:non-branded queries only. Downsampling/Oversamplingof dataset Random forest balanced : 50 decision trees K-Folds CrossValidation 80/20 Features:linking metrics ranking inside of the SERP for URL an domain
  36. 36. RESULTS
  37. 37. SUCCESS RATES top1 : 98,25% top3 : 95,27% top5 : 92,35% top10 : 85,3% (between us, that’s neat)
  38. 38. SO, WHAT IS INSIDE OURBOX?!
  39. 39. FEATURES WEIGHTS IN OURMODEL
  40. 40. FEATURES WEIGHTS IN OURMODEL
  41. 41. How dowe comparevs randomness? RANDOM NOT RANDOM
  42. 42. RANDOM NOT RANDOM Howdo we compare vs randomness?
  43. 43. TALKINGABOUT Contentstrategy… withoutMENTIONINGcontent?! • “Content is king” : yeah, but no. • Our model worked without anything related to: – Speed score, – content length, – design, etc. • BUT our keyword sample is quite specific
  44. 44. Key takeaways • Ranking is overtaking your competitors, monitoring your environement is crucial. • build EARN backlinks directly to your URLs • Get “juice” to your URLs with inner linking • Host your message on big authority websites
  45. 45. Use yourinternallinkingwisely
  46. 46. Add someexternallinking,too
  47. 47. teşekkür ederim k@seobserver.com @512banque

×