Impressions on a given search term can often take multiple routes through an account, and those routes are not equally valuable.
In search campaigns, each of those routes is a keyword.
A keyword is best understood as an offer to buy ad inventory against a query. Exact match is closer to an unconditional offer, to buy all instances of that query and its neighbours, provided the budget and bidding requirements are met. A broad match keyword is closer to a conditional offer made across a much wider range of queries, to buy that inventory when the signals are right.
Two things follow from the kind of offer you make:
- Which instances of a query you compete for at all
- Where the ones you win end up (the ad, the landing page, the bid)
Call them selection and routing. Both decide what a search term is worth to you, and the second gets a good deal less attention than the first.
One search term, thirteen routes
Deciding which route a search term takes through an account has become a larger part of the job as the keyword-to-search-term relationship has loosened (and it has loosened a great deal over the last few years).
Here is a relatively extreme example. Over four months, the single search term “energy certificate uk” reached one account through thirteen different keywords, sitting in six ad groups across two campaigns, with markedly different results.
| Route (keyword) | Share of spend | CVR | CPA index |
|---|---|---|---|
| epc certificate | 6% | 18.8% | 34 |
| how do i get an energy performance certificate | 6% | 17.8% | 35 |
| get epc certificate | 16% | 8.0% | 94 |
| energy performance certificate | 29% | 5.7% | 99 |
| energy certificates uk | 9% | 6.7% | 145 |
| energy certificate uk | 25% | 1.9% | 276 |
| energy proficiency certificate | 6% | 0.0% | no conversions |
Cost per conversion indexed against 100 for the search term as a whole. Six further keywords took the remaining 3% of spend between them and returned nothing.
The keyword that is the search term itself takes 25% of the spend and returns 9% of the conversions, at a CPA index of 276 against 100 for the term as a whole. Its conversion rate was 1.93% against 8.32% for every other route combined. The exact match keyword was, simply, the least efficient way for this search term to enter the account.
It also had nearly the lowest CPC in the table, so the damage was done by conversion rate alone.
There are two plausible reasons for that, and both may well be at work.
The keyword was making an unconditional offer, doing nothing to discriminate between the auctions it could enter (selection). It also sat in a miscellaneous ad group with nothing particular to say to anybody (routing). Either mechanism would produce a number like that one, and this account can’t separate them.
Why this happens
An unconditional offer puts your ad forward for every instance of the query string. A conditional one does so only where the accompanying signals suggest that the search is worth competing for. The query string is identical either way, but the set of searches in which it appears is not, and neither is the conversion rate.
The auctions a keyword makes you eligible for aren’t equally attractive. Each has its own chance of converting and its own price to win, which together give it an expected cost per conversion, and those run from cheap through to expensive. The bidding system works its way up that spread (in theory) until it reaches the target.
A conditional offer covers a wide spread, so the system can confine itself to the good end of it. An unconditional offer on a given search query covers a narrow spread, and the system has to take most of what is there, expensive end included.
This is the efficiency side of the argument Google has been making for broad match for years, and the logic behind the recommendation that still appears in the interface, promising “more conversions at a similar or better ROI” from expanding the reach of your Smart Bidding campaigns.
Removing the direct route
The same campaign, a few weeks later. Two exact match keywords that had held a monopoly on two query strings were paused on 22 May, and one of them appears in the table above.
Over the following 52 days, compared with the 52 days before, clicks on those strings fell by 26% and spend by 28%, while conversions rose by 178% and cost per conversion fell by 74%. The number of routes carrying the two strings went from two to ten.
Both mechanisms are visible here. Volume genuinely fell, so some instances were being declined rather than rerouted (selection). The instances that survived landed across five different ad groups instead of one (routing).
Accepting a 26% drop in volume on a query that you know converts is a more uncomfortable decision than rerouting its impressions, but it did much of the good work here.
None of this involved AI Max. It was an old-fashioned exact-versus-broad phenomenon. The same mechanism, though, can run at a considerably greater scale when automation joins the party.
The automated version
A different account this time, in the same line of business. The campaign had four active ad groups: one containing a core keyword in both broad and exact, which acted as the campaign’s wide-net matching engine while being its worst performer on CPA, and three themed ad groups covering location, general and near-me searches.
AI Max search term matching went live on 20 May, with text customisation following on 1 July. That leaves 20 May to 30 June as a cleanly analysable ‘after’ period.
| Ad group | Share of spend, before | Share of spend, after | CPA before | CPA after |
|---|---|---|---|---|
| Core keyword (broad + exact) | 46.6% | 3.3% | £50.28 | £72.91* |
| Location | 19.4% | 46.2% | £32.90 | £43.00 |
| General | 25.3% | 24.1% | £49.68 | £30.18 |
| Near me | 8.6% | 26.4% | £27.95 | £34.68 |
*The core ad group took four conversions in the after period, so its CPA there isn’t a figure to lean on.
On a daily spend around 17% higher, conversions per day rose by about a third and cost per conversion fell from £42.87 to £37.33.
OK, a relatively rare case of AI Max taking a clear win. But a dig into the search terms makes the mechanism for this improvement particularly interesting.
The outlying queries the core keyword had been catching in April – certificate register lookups, postcode checkers, “how to get” queries, the name of an accreditation body – were still arriving in July. They had simply stopped arriving through the core keyword. The same strings now turn up in the location, general and near-me ad groups, under the AI Max search term match type, and what changed for them was the ad they were shown and the price paid per click (the landing page was shared between them).
Look down the last two columns, though, and three of the four ad groups came out worse than they went in. Only the general ad group improved. The campaign improved anyway, because nearly half of its spend left the ad group that had been running at £50.28 and moved onto better ones. A blended figure that improves while most of its components deteriorate flatters whoever is reporting it, so the gain here is largely a mix effect and it is worth saying so.
In a recent piece well worth reading on evaluating AI Max, Brad Geddes warns that expanded matches may include traffic you were already receiving from your exact and phrase keywords, and that AI Max will divert spend away from your handpicked keywords if budget is tight. Both are sensible checks to run.
While that diversion is presented as a hazard to watch out for, it was exactly the mechanism behind the improvement in this campaign, with the core keyword going from 46.6% of spend to 3.3% without being touched. Rerouting can be a benefit, provided there are potentially better paths for those searches to take.
We are used to thinking of Google’s automated systems as helping themselves to low-hanging fruit that was already on its way to a perfectly good destination. It is worth being aware that even a deliberate, well-built structure can produce routings that are not the most efficient available, and that a bit of automated shaking of the tree can sometimes be beneficial.
Only one of the two mechanisms is on show here, though. AI Max moved these queries into better ad groups, and there is nothing in the search terms to suggest it declined any of them along the way. Selection stayed roughly where it was. What improved was the routing.
What makes a route worth taking
Selection and routing are separate levers.
Pausing a keyword, as in the second example, can reach both at once, declining some impressions outright and sending others down a more favourable path. In the AI Max campaign only the second lever moved, with the matching layer redistributing thousands of queries at a stroke while turning none of them away. Neither approach is inherently the more powerful of the two.
What decides how much either is worth is whether the account has anywhere better to send a query.
This doesn’t have to mean separate landing pages. In the AI Max campaign the four ad groups used the same landing page. What separated them was ad text, with the location ad group leaning heavily on DKI and the others emphasising different messaging.
So the prerequisite for routing to be worth anything is a set of meaningfully differentiated ad groups, which is a considerably lower bar than a set of separate page builds. Before turning on AI Max search term matching, or broadening match types to the same end, the question worth asking is whether the campaign has anywhere to send a query other than where it already goes.
For all the push towards consolidation that accompanies Google’s drive towards automation, a single campaign with a single ad group provides only one route. This particular, and perhaps unintended, benefit of automation depends on having distinct ad groups available to act as different routes.
