$30 off During Our Annual Pro Sale. View Details »

Oh My MUM Strikes Back - Or how to win the battle of Generative Search

Oh My MUM Strikes Back - Or how to win the battle of Generative Search

Generative Search may be scaring... and we already see AI gurus declaring the (again) soon to be death of SEO and Search Marketing.
But the reality is not such.
In this deck you will see how you must rediscover classic concepts like taxonomy and ontology to be sure you'll survive the AI revolution in Search.

gianluca fiorelli

June 19, 2023
Tweet

More Decks by gianluca fiorelli

Other Decks in Marketing & SEO

Transcript

  1. Gianluca Fiorelli
    International & Strategic SEO Consultant
    Website: https://www.iloveseo.net
    Twitter: @gfiorelli1
    Chocolate and Movie lover.
    Painting minis expert.
    Italian adopted by Spain, and citizen of the
    world.

    View Slide

  2. View Slide

  3. OH, MY MUM!
    Strikes Back

    View Slide

  4. Universal
    Search

    View Slide

  5. Search
    Features

    View Slide

  6. More Search
    Features, E-E-A-T
    et al

    View Slide

  7. More Search
    Features, E-E-A-T,
    et al, and
    SGE

    View Slide

  8. The truth is that we
    shouldn’t despair
    for the future

    View Slide

  9. As long as we think
    strategically and not
    just in a reactive way.

    View Slide

  10. Occam's Razor:
    The goal of any SEO strategy
    is to bring as much qualified
    organic traffic to the right
    pages of our website to
    convert it into customers.

    View Slide

  11. But if before these users had
    a clear path to reach the
    desired destination...

    View Slide

  12. … now they can take very
    different paths to reach that
    same destination.

    View Slide

  13. And with the
    future arrival of
    SGE, these paths
    are going to be
    transformed into a
    true labyrinth.

    View Slide

  14. ‘’Gattopardismo’’:
    Everything must change so that
    everything continues as before.

    View Slide

  15. View Slide

  16. In the case of
    Google, the
    integration of SGE
    is nothing more
    than another step
    in delving into the
    concept of the
    Messy Middle.

    View Slide

  17. However, the
    evolution of
    generative "AI"
    and LLMs presents
    us with some
    challenges and
    some (great)
    opportunities.

    View Slide

  18. The biggest
    challenge is to
    understand this
    evolution, to be
    aware that it is
    accelerated and that
    for this reason our
    strategy must be
    based on reliable
    constants in the
    medium/long term.

    View Slide

  19. At the same time, however,
    we have to take advantage of
    this same evolution to ensure
    that our strategy based on
    reliable constants in the
    medium and long term is not
    just a one-day success.

    View Slide

  20. BUT

    View Slide

  21. One does not simply
    enter into Mordor

    View Slide

  22. all of this has happened before.

    View Slide

  23. all of this has happened before.
    and all of this will happen again

    View Slide

  24. But in a new context.

    View Slide

  25. With respect to the past year, some things
    have changed and went partly unnoticed.
    One of the most relevant changes is
    related to the Product Knowledge Graph.
    Here we see how it was presented in the
    summer of 2022 for “Star Wars: Legion
    minis”.

    View Slide

  26. Google
    Merchant
    Google
    Manufacturer
    Insights
    based on
    product
    taxonomies
    Reviews.
    Analysis-based
    Q&A
    Expert Reviews
    Videos
    2023

    View Slide

  27. Google
    Merchant
    Google
    Manufacturer

    View Slide

  28. Insights
    based on
    product
    taxonomies
    Expert Reviews
    Outreach actions are more necessary
    than ever, and not just to get links
    (which are still important), and they
    should even be preventive.
    Analyze who "speaks" about your
    competitors and is used by Google as a
    source of EXPERT information.

    View Slide

  29. Expert Reviews
    Outreach actions are more necessary
    than ever, and not just to get links
    (which are still important), and they
    should even be preventive.
    Analyze who "speaks" about your
    competitors and is used by Google as a
    source of EXPERT information.
    Top Consideration and Critic Review are
    not conditioned by QDF.
    Insights
    based on
    product
    taxonomies

    View Slide

  30. Expert Reviews
    Outreach actions are more necessary
    than ever, and not just to get links
    (which are still important), and they
    should even be preventive.
    Analyze who "speaks" about your
    competitors and is used by Google as a
    source of EXPERT information.
    Top Consideration and Critic Review are
    not conditioned by QDF.
    In some cases, Google uses comments
    present in the cited source.
    Insights
    based on
    product
    taxonomies

    View Slide

  31. Expert Reviews
    Outreach actions are more necessary
    than ever, and not just to get links
    (which are still important), and they
    should even be preventive.
    Analyze who "speaks" about your
    competitors and is used by Google as a
    source of EXPERT information.
    Top Consideration and Critic Review are
    not conditioned by QDF.
    In some cases, Google uses comments
    present in the cited source.
    The taxonomy depends on the ontology
    that we have defined for the product.
    Insights
    based on
    product
    taxonomies

    View Slide

  32. Reviews.
    Analysis-based
    Q&A
    The reviews can be both from "Google", that is, extracted from the
    reviews made directly on Google Merchant (paid and/or organic), but
    they can also be reviews indexed from websites.

    View Slide

  33. Questions in evidence appear only in desktop searches and are
    generated through what seems to be NLP analysis of the reviews
    themselves.
    What implicit questions does each review answer?
    What are the reviews that can be grouped under the same question?
    What are the most relevant and common questions?
    Reviews.
    Analysis-based
    Q&A

    View Slide

  34. With ChatGPT or OpenAI APIs and embeddings we can do something
    similar ourselves using our reviews (or others' or public sources) to be
    able to identify pain points and classify them by clusters and semantic
    neighborhood.
    For that we will have to prepare the dataset with Python:
    Reviews.
    Analysis-based
    Q&A

    View Slide

  35. Then, we cluster with K-Means :
    Finally, we can take advantage of GPT-4 to get an automated
    description of each cluster:
    1. We extract some reviews from each cluster,
    2. we use them to compose a prompt that asks GPT-4 to analyze the
    similarity and
    3. to generate a title or description accordingly.
    4. Generate the questions to which they implicitly respond.
    Reviews.
    Analysis-based
    Q&A

    View Slide

  36. Videos
    Again, outreach actions are more
    necessary than ever, and not just to get
    links (which are still important), and
    should even be preventive.

    View Slide

  37. Videos
    Again, outreach actions are more
    necessary than ever, and not just to get
    links (which are still important), and
    should even be preventive.
    Note that the mobile SERPs also feature
    Shorts.

    View Slide

  38. Videos
    Again, outreach actions are more
    necessary than ever, and not just to get
    links (which are still important), and
    should even be preventive.
    Note that the mobile SERPs also feature
    Shorts.
    Analyze who "speaks" about your
    competitors and is being used as a
    source of EXPERT information.

    View Slide

  39. Videos
    Pay attention to the evolution of
    YouTube and the Creator Economy,
    because it too is accelerated.

    View Slide

  40. Videos
    And remember that up to 45% of users watch YouTube on
    television (USA data).
    Therefore: 4K, high quality and consistent Thumbnails,
    Shorts, introductions with the "Best of" of what we are going
    to see, be careful with the "Series" function for Playlists and
    their titles.

    View Slide

  41. The battle of
    Hoth

    View Slide

  42. SGE is in beta and in
    many cases this is
    what it offers us.
    This does not mean
    that you should not
    study, quite the
    contrary.

    View Slide

  43. Answer generated by
    Google
    Sources used to
    generate the answer
    Shopping
    Graph
    Los rankings
    orgánicos no
    aseguran la cita
    Shopping Graph
    Button to continue
    the conversation
    "Related Conversations". They
    are not the same as PAAs

    View Slide

  44. We can expand the conversation to see
    which part of the answer is based on one
    source or another.
    If the query is a question, in the organic
    results Google shows the answer, which
    makes us suspect the existence of
    "response indexing".

    View Slide

  45. If we click on a recommended product, the Shopping Graph tab opens in
    the same SERP.
    The description we see can be taken either from the description made by
    the Manufacturer or from a review, as in this example.

    View Slide

  46. Generative AI
    and semantics

    View Slide

  47. The Force is calling
    Just let it in

    View Slide

  48. "For every parent,
    their child is the most
    beautiful in the
    world." The same
    goes for websites.

    View Slide

  49. Unstructured Structured

    View Slide

  50. T
    AXONOMY

    View Slide

  51. A taxonomy is a method of organizing data.
    Taxonomy represents the formal structure of
    classes or types of objects within a domain.
    A taxonomy can also be a set of chosen terms used
    to retrieve content online (for example, a
    website's taxonomy).

    View Slide

  52. View Slide

  53. Alderaan
    Appearances (Episode IV)
    Affiliation (Galactic Empire)
    Locations (Aldera
    Spaceport)
    Terrain (Mountains)
    Vehicles (Tantive IV)
    Videos (“Destruction of
    Alderaan)
    Gallery
    History
    Related (we will see this in
    a few slides)
    Locations

    View Slide

  54. View Slide

  55. View Slide

  56. But as good as the
    taxonomy we've
    created is, is it
    consistent with
    potential Search
    Journeys within
    Google?

    View Slide

  57. Contextual Search Menu Filters

    View Slide

  58. Images Search tags
    Level 0
    Level 1
    Level 2
    Level 3
    Level 4

    View Slide

  59. Images Search tags con
    Topically.io

    View Slide

  60. People Also Ask
    Don't use PAAs just to get new content ideas.
    Use them to create thematic clusters and, above all,
    second and third level clusters,
    and to optimize their internal linking.

    View Slide

  61. SGE Related Questions
    Collect responses and do entity analysis
    with embeddings to get new data
    to integrate with what you’ve
    already retrieved.

    View Slide

  62. AI Related Questions with Perplexity.ai
    Collect responses and do entity analysis
    with embeddings to get new data
    to integrate with what you’ve
    already retrieved.
    Save the sources, and analyze them.

    View Slide

  63. AI Related Questions with Perplexity.ai
    Dig into the ”related questions” as you
    do in the case of People Also Ask.
    Use them to explore topical clusters
    possibilities.

    View Slide

  64. USE WORDLIFT ADD-ON FOR SHEET TO ANALYZE THE ENTITIES TARGETED BY THE PAGES RANKING IN TOP 10

    View Slide

  65. S
    T
    A
    R
    W
    A
    R
    S
    L
    E
    G
    I
    O
    N
    TABLETOP GAME
    TERRAIN PAD
    MINIS
    ARMIES
    HEROES
    CARDS
    RULES
    REBELS
    STORMTROOPERS
    LUKE SKYWALKER
    DARTH VADER
    FIGURE
    ASSEMBLING
    AND
    PAINTING
    GUIDES
    EXPERTS
    PAINTS
    ATOMIC MASS
    GAMES
    (MANUFACTURER)
    ’’PRODUCT’’
    CRITICAL REVIEWS
    BLOGS
    MAGAZINES
    VIDEOS
    SHOPPING GRAPH MERCHANTS UGC REVIEWS
    EXPERTS
    JOURNALISTS
    CREATORS
    3D PRINTED
    Finally, update your
    architecture

    View Slide

  66. Use Google Trends to keep the taxonomy up to date or to change it based on seasonality.

    View Slide

  67. Speed up your Trend analysis work with a tool like Keytrends
    Speed up your Trend analysis work with a tool like Keytrends

    View Slide

  68. ONTOLOGY

    View Slide

  69. Ontology is a formal representation of knowledge
    about a domain, including the concepts within
    that domain and the relationships between those
    concepts.

    View Slide

  70. An ontology provides schema information and
    sometimes axioms or rules for consistency.
    Rather, taxonomies provide information about
    hierarchies of concepts and things.

    View Slide

  71. Alderaan
    Appearances (Episode IV)
    Affiliation (Galactic Empire)
    Locations (Aldera
    Spaceport)
    Terrain (Mountains)
    Vehicles (Tantive IV)
    Videos (“Destruction of
    Alderaan)
    Gallery
    History
    Related
    Locations

    View Slide

  72. View Slide

  73. Alderaan
    Base Espacial de
    Alderaan
    Palacio Real de
    Alderaan
    Bail Organa
    Reina Breha
    Organa
    Leia Organa
    Alianza Rebele
    Crucero
    alderano Tantive
    IV
    Imperio
    Galáctico
    Grand Moff
    Tarkin
    Muerte Negra
    Obi-Wan Kenobi

    View Slide

  74. The Alderaan Royal Palace is the seat of the Alderaan government
    headed by Bail Organa, a senator of the Republic and the Galactic
    Empire and one of the most prominent members of the Rebellion.
    Married to Queen Breha Organa, the two adopted the daughter of
    Padme Amidala and Anakin Skywalker: Leia Organa.
    In her childhood, Leia Organa, heir to Alderaan, was rescued by Obi-
    Wan Kenobi and, once they assumed diplomatic roles, was key to the
    theft of the Death Star's secret plans thanks to the versatility of her
    Alderaan cruiser.
    However, she was captured by Grand Moff Tarkin, who destroyed the
    planet Alderaan with the Death Star.
    Prior to her capture, Leia Organa managed to send a distress message
    to Obi-Wan Kenobi via an astromechanic droid.

    View Slide

  75. schema

    View Slide

  76. Markup for Hotels
    https://schema.org/docs/hotels.html basado en Accommodation
    Ontology Language Reference

    View Slide

  77. Markup for automative

    View Slide

  78. Markup for banks and financial institutions

    View Slide

  79. SameAs era importante. Ahora todavía más

    View Slide

  80. CONTENT

    View Slide

  81. ENTITY SALIENCE and CONTEXT
    Surviving the fall of the Galactic Empire, Moff Gideon soon takes
    command of a massive Imperial force in disarray, reorganizing it to
    counter the newly established New Republic.
    Armed with the Mandalorian black lightsaber, Moff Gideon will
    find his fiercest opponents in Mando, Grogu, and the
    Mandalorians led by Bo-Katan Kryze.
    An entity represents a phrase in text
    that is a known entity, such as a
    person, organization, or location.
    An entity’s salience score provides
    information about the importance or
    centrality of an entity to all text in the
    document.
    Scores close to 0 are less prominent,
    while scores close to 1.0 are highly
    prominent.

    View Slide

  82. Moff Gideon appears in all three seasons of The Mandalorian that
    have aired so far.
    Played by cult actor Giancarlo Esposito (Breaking Bad), Moff
    Gideon has been the subject of many theories by fans of the
    series, such as his role in cloning Emperor Palpatine or his own
    death at the end of the third season.
    Surviving the fall of the Galactic Empire, Moff Gideon soon takes
    command of a massive Imperial force in disarray, reorganizing it to
    counter the newly established New Republic.
    Armed with the Mandalorian black lightsaber, Moff Gideon will
    find his fiercest opponents in Mando, Grogu, and the
    Mandalorians led by Bo-Katan Kryze.
    ENTITY SALIENCE and CONTEXT

    View Slide

  83. Moff Gideon, like all Star Wars: Legion minis, comes in hard plastic.
    Easy to assemble, it faithfully reproduces the face of the actor
    Giancarlo Esposito who played him in The Mandalorian.
    The mini is accompanied by 1 unit card, 1 upgrade card, 3
    command cards, 1 token, and 1 instruction sheet.
    The Moff Gideon miniature comes not painted.
    Surviving the fall of the Galactic Empire, Moff Gideon soon takes
    command of a massive Imperial force in disarray, reorganizing it to
    counter the newly established New Republic.
    Armed with the Mandalorian black lightsaber, Moff Gideon will
    find his fiercest opponents in Mando, Grogu, and the
    Mandalorians led by Bo-Katan Kryze.
    ENTITY SALIENCE and CONTEXT

    View Slide

  84. Entity analysis through embedding for Lens
    With Lens we know that people
    can search for images.
    Thanks to embeddings, visual
    components are transformed into
    numbers.
    Thanks to our previous work, we
    can easily see the neighborhood
    between products also if they are
    taxonomically different.

    View Slide

  85. Lens works for objects, so if it
    recognizes them it highlights them so
    they can be selected.
    1. Embeddings
    2. Entity clustering
    3. neighborhood analysis
    4. Optimizing images for “related
    products” in Lens
    (note: this process can be used equally
    for related products at the website
    level).
    Entity analysis through embedding for Lens

    View Slide

  86. In pink
    Entity analysis through embedding for
    Multisearch

    View Slide

  87. AI Images to cover stock shortages

    View Slide

  88. AI to improve stock image quality
    We can use Super-Resolution for images, which
    can enlarge and enhance images using
    generative AI and Deep Learning.
    A good model is ESRGAN (Enhanced Super
    Resolution Generative Adversarial Network)
    available on Tensorflow Hub (h/t Andrea Volpini
    aka @cyberdandy):
    https://www.tensorflow.org/lite/examples/super_resolution/overview
    https://www.tensorflow.org/hub/tutorials/image_enhancing
    https://arxiv.org/abs/1809.00219
    https://github.com/hiram64/ESRGAN-tensorflow

    View Slide

  89. Store window display of a colorful
    collection of Fender electric guitars::3
    with vibrant colors and exquisite details::2
    Canon EOS 5D Mark IV, 24-70mm lens, f/4,
    ISO 100, 1/100 sec
    Textures and details.
    Warm lighting and city reflections, high-
    end fashion, urban setting
    --ar 1:1 --v 5.1 --style raw --s 1000 --q 2
    --no people, watermarks
    seed 901509742
    And, do not prompt things like “Imagine a unicorn riding in the clouds”. Be precise and specific!

    View Slide

  90. Store window display of a colorful
    collection of Fender electric guitars::3
    (things that we want to create and relative
    weight of this part of the prompt)
    And, do not prompt things like “Imagine a unicorn riding in the clouds”. Be precise and specific!

    View Slide

  91. with vibrant colors and exquisite details::2
    (define the overall style and weight it)
    And, do not prompt things like “Imagine a unicorn riding in the clouds”. Be precise and specific!

    View Slide

  92. Canon EOS 5D Mark IV, 24-70mm lens, f/4,
    ISO 100, 1/100 sec
    (indicate the tech you want the AI to
    simulate for you)
    And, do not prompt things like “Imagine a unicorn riding in the clouds”. Be precise and specific!

    View Slide

  93. Textures and details
    (do not forget to indicate the materiality of
    the image)
    And, do not prompt things like “Imagine a unicorn riding in the clouds”. Be precise and specific!

    View Slide

  94. Warm lighting and city reflections, high-
    end fashion, urban setting
    (be precise when coming to environmental
    or studio lighting)
    And, do not prompt things like “Imagine a unicorn riding in the clouds”. Be precise and specific!

    View Slide

  95. --ar 1:1 --v 5.1 --style raw --s 1000 --q 2
    --no people, watermarks
    (learn how to use the parameters, and
    make us of negative weighting for
    obtaining clean images)
    And, do not prompt things like “Imagine a unicorn riding in the clouds”. Be precise and specific!

    View Slide

  96. seed 901509742
    (discover the function of “seed” to
    maintain consistency between generated
    images).
    And, do not prompt things like “Imagine a unicorn riding in the clouds”. Be precise and specific!

    View Slide

  97. Because, using
    prompts as a work
    tool, if we are not
    very precise, we will
    run the risk of
    working from the
    wrong bases and,
    thus, reaching
    disastrous decisions.

    View Slide

  98. Happy AI…
    and may the Force
    be with you.

    View Slide