g_i = Google's identifier for customer i. s_j = Merchant identifer for customer j.
If g_i == s_j then customer i == j. These identifiers might be phone numbers or email addresses or other identifiers that both parties have.
Neither party wants the other to learn any identifiers it doesn't already known about.
So Google picks a random secret value G and sends g_i^G to the Merchant. The Merchant picks random secret value T and sends g_i^(G* T) back to Google. Additionally the merchant sends s_j^T to Google.
Then Google calculates s_j^(T* G). If s_j^(T* G) == g_i^(G* T) then s_j == g_i and i == j. So know Google knows the exact set of their users who purchased something at the Merchant.
Additionally, for each s_j^(T), the Merchant sends a homomorphically encrypted value for the amount they spent. Then Google can perform a homomorphic addition on these encrypted values of only the intersection it calculated. The Merchant can then decrypt it to get the total sum and share it back to Google.
So in the scheme I described
1) Google still learns who is purchasing at which merchant. 2) Google does not learn individual amounts. 3) Merchants can't perform the same calculation to learn which users saw their ads unless Google sends s_j^(T* G) back to them (not pictured in the slides).
Furthermore, although reversing a one-way hash is problematic, computing hashes for known inputs (such as an email address) is straight-forward. So it is possible to just test or probe datasets for known identifiers and persons. The potential for hash collisions is of little comfort.
In short, they're able to compute the intersection of the set of users who have viewed an ad with the set of people who purchased a product in a store, without either party disclosing their side of the set.