Martech Partners Contact
MartechPartners

BigQuery Optimized Mode for AI.IF and AI.CLASSIFY, Plus AI.RANK

Google Cloud · Editorial Team · 18 November 2026

BigQuery's release notes for 17 November 2026 mark four features as generally available: optimized mode for the AI.IF and AI.CLASSIFY functions, AI.RANK, a rerank option in AI.SEARCH, and REST API control over cross-region dataset replication and failover reservations. The first three change how you run language-model work inside SQL. The fourth changes how you manage disaster recovery for datasets.

If your team has been running LLM-backed classification or filtering over millions of rows, the cost and latency angle matters most. If you own resilience for a warehouse that the business depends on, the dataset API change is the one to read twice.

Key takeaways

  • Optimized mode for AI.IF and AI.CLASSIFY is generally available and exists to cut token consumption and query latency on large datasets.
  • AI.CLASSIFY in optimized mode now handles multi-label classification, and more types of queries are in scope for optimized mode.
  • The Job information tab in the Query results pane reports the relative quality of optimized results, so you can check what you traded for speed.
  • AI.RANK and the rerank option in AI.SEARCH both reorder results by how well text answers a query. Both are generally available.
  • datasets.insert, datasets.patch, datasets.update and datasets.get can now configure and monitor cross-region replication and attach or detach failover reservations.
  • Pilot each change on a small, labelled sample first. Roll out by workload, not all at once.

What changed in the 17 November 2026 BigQuery release

The notes for 17 November 2026 list four items, and all four are described as generally available.

First, optimized mode for the managed AI functions AI.IF and AI.CLASSIFY. The stated goal is to reduce LLM token consumption and query latency when processing large datasets. The notes add three details: more types of queries are in scope for optimized mode, AI.CLASSIFY supports multi-label classification in this mode, and the Job information tab in the Query results pane shows information about the relative quality of the optimized results.

Second, the BigQuery REST API methods datasets.insert, datasets.patch, datasets.update and datasets.get can configure and monitor cross-region dataset replication. They can also attach or detach failover reservations for managed disaster recovery.

Third, the rerank option in AI.SEARCH. It reorders search results by how well the text in the column being searched answers the query.

Fourth, the AI.RANK function. It reorders the rows of a table by how well the text in a specified column answers a search query.

Here is the whole release at a glance.

Change Who it affects First action
Optimized mode for AI.IF and AI.CLASSIFY Teams running LLM filtering or labelling over large tables Benchmark on a labelled sample, read the quality information in the Job information tab
Multi-label classification in AI.CLASSIFY optimized mode Teams tagging rows with more than one category Review whether one-label-per-row workarounds can be simplified
AI.SEARCH rerank option Teams building search or retrieval over text columns Compare ordering with and without rerank on real queries
AI.RANK Teams that already have candidate rows and need relevance ordering Test on a known set with agreed best answers
REST API for replication and failover reservations Platform, data engineering and resilience owners Capture dataset replication settings as code and review them

Should you switch AI.IF and AI.CLASSIFY to optimized mode?

Probably, for the big, repetitive jobs. Not blindly.

The case for optimized mode is simple. A function that asks a language model a question about every row gets expensive and slow as the table grows. Google says optimized mode is meant to reduce token consumption and latency on large datasets. Picture a content team classifying a few million product descriptions or support tickets by topic. That is exactly the shape of work where per-row model calls hurt.

The case for caution is in the notes too. The Job information tab shows relative quality of the optimized results. Google would not need to surface that if optimized output were identical to a standard run in every case. Treat it as a speed-and-cost setting with a quality dimension you have to look at.

A sensible decision rule:

  1. Switch first where the question is narrow and the labels are few, such as a yes or no filter or a short category list.
  2. Keep standard behaviour for decisions with a real cost of error, like routing legal or compliance text, until your sample says optimized mode is good enough.
  3. Do not move a workload on the strength of one query. Run it on a representative sample of rows, including awkward ones.

The notes do not give numbers for savings, so don't promise any internally. Measure your own.

How to test optimized mode before it touches production

Build a small evaluation set. A few hundred rows is usually enough to see a pattern, and the right number depends on how varied your text is. Have someone who knows the domain label them by hand. A marketing analyst who knows what counts as a 'pricing complaint' beats a generic guess.

Then run the same queries two ways and compare.

  • Agreement with your hand labels, row by row, for both runs.
  • Bytes and slot time for each job, plus the job duration, from the job history you already have.
  • The relative quality information in the Job information tab for the optimized run.
  • Disagreements, read by a human. Counting them is not enough. Look at which kinds of rows fall out.

One trap: sample only easy rows and every setting looks great. Include the ambiguous middle, short texts, mixed-language text and anything with sarcasm or negation.

Once you have a result, write the acceptance bar down before you look at production. For example, optimized mode must stay within an agreed margin of the standard run on the labelled set, and job time must drop meaningfully. The margin is a business call, not an engineering default.

What multi-label classification in AI.CLASSIFY changes

Many real tagging jobs aren't single-answer. A customer review can be about delivery and price. A sales call note can mention a competitor and a renewal risk. The notes say optimized mode supports multi-label classification with AI.CLASSIFY, so one row can carry more than one category.

That opens a few practical options, and also a few new decisions.

If you have been running a classifier once per category, or forcing a single best label and losing information, review that design. A multi-label pass may simplify your pipeline. Check the Google documentation for the exact syntax and output shape before you rewrite anything, because downstream tables and dashboards will depend on it.

The new decision is how you count. A single-label table sums to the number of rows. A multi-label table doesn't. Reports that show 'share of rows by topic' need a clear definition, or somebody will add the percentages up and get more than 100.

Also agree on what 'no label applies' means in your taxonomy. Without a rule, the same row gets treated differently by different analysts.

AI.RANK and the AI.SEARCH rerank option: where they fit

These two features share one idea. Given a query and some text, put the best answers first.

AI.RANK reorders the rows of a table by how well the text in a column you specify answers a search query. The rerank option in AI.SEARCH does the same reordering on search results. The practical difference is where you start. With AI.SEARCH you are searching and can ask for the reordering as part of that. With AI.RANK you already have rows, perhaps from an earlier filter or join, and you want them ordered by relevance to a question.

Where teams will use them:

  • Internal knowledge search over documentation, tickets or call summaries held in BigQuery.
  • Retrieval for an assistant that answers questions from warehouse text, where the order of the top results decides what the model sees.
  • Analyst workflows, such as pulling the most relevant past campaign briefs for a new brief.

The notes don't state limits, pricing or latency for either feature, so don't assume anything. Test on the questions your users really ask, and compare the order against what a subject-matter expert would pick. A search tool that looks clever on ten demo queries can still fail on the eleventh.

What to leave alone: if your search results are already good and nobody complains, reranking is not urgent. Add it where relevance is the visible weakness.

Managed disaster recovery now has a REST API path

The fourth change is quieter and matters more to platform owners than to analysts. The datasets.insert, datasets.patch, datasets.update and datasets.get methods can now configure and monitor cross-region dataset replication, and attach or detach failover reservations for managed disaster recovery.

What that gives you is a way to treat replication and failover settings as part of your infrastructure definition. You can read them with datasets.get, change them with the patch or update methods, and set them at creation with datasets.insert. Because these are API calls, they can sit in scripts and pipelines that your team already reviews.

A few things worth doing:

  1. List which datasets are business-critical: attribution tables, customer data, finance reporting. Decide which of them need cross-region replication at all.
  2. Record the intended replication and failover reservation settings for each one in your infrastructure code or runbook.
  3. Use datasets.get on a schedule to check that what is deployed matches what you intended, and alert on drift.
  4. Decide who is allowed to attach or detach failover reservations. That action changes how a failover behaves, so it needs an owner and an approval path.

The notes say nothing about pricing, quotas or the exact fields involved. Read Google's disaster recovery documentation before you design around it, and confirm cost with your Google Cloud account team. Replicating data across regions is rarely free in any cloud, and reservations are a capacity commitment, so budget before you switch anything on.

A failover setting nobody has tested is a hope, not a plan. Schedule a drill in a non-production project and write down what the team had to do by hand.

Who should act, and a rollout plan

Not everyone needs to do everything. Here's a split that works for most teams.

Step Owner Done when
Pick two or three high-volume AI.IF or AI.CLASSIFY workloads Analytics lead Workloads chosen and a labelled sample built
Run standard and optimized mode side by side Data engineer Agreement, time and cost recorded
Review quality information and disagreements Domain expert Written go or no-go per workload
Pilot AI.RANK or AI.SEARCH rerank on real queries Search or product owner Ordering compared with expert picks
Capture replication and failover reservation settings via the REST API Platform engineer Settings stored as code, drift check running
Run a failover drill outside production Platform and data owners Runbook updated with what went wrong

Roll out by workload. Move the lowest-risk job first, watch it for a couple of weeks, then continue. Keep a one-line way back to the standard behaviour for every job you move.

How to measure whether it worked

Pick a handful of numbers before you start, and keep them boring.

For AI functions: job duration, bytes processed or slot time from your normal job history, and agreement with human labels on a fixed sample that you re-run monthly. Add the relative quality information the Job information tab shows, and note it beside each run so you can see if it moves.

For ranking and search: how often the answer an expert would choose lands in the top few results, measured on a saved list of real questions. Keep that list. It becomes your regression test the next time anything changes.

For disaster recovery: whether the deployed settings match intent at each check, and how long your last drill took from decision to working reports.

If a metric can't be read from data you already collect, ask whether you really need it.

What not to do

Don't migrate everything because the feature is generally available. General availability tells you Google considers the feature ready for production use. It doesn't tell you it is right for your data.

Don't skip the labelled sample. Cost numbers without a quality check are half a result.

Don't expose multi-label output to dashboards without defining how percentages are calculated.

Don't change failover reservations as part of a routine data change. Keep that behind its own approval.

Don't treat the new REST API support as a replacement for a tested recovery plan. It makes configuration repeatable. It doesn't prove the plan works.

And don't rewrite working pipelines purely to use the newest function names. If a job is cheap, quick and accurate, leave it.

Where an implementation partner helps, and where you can go alone

Much of this you can do yourself. A data engineer with a free afternoon can run an optimized-versus-standard comparison on a sample. An analyst can try AI.RANK on a saved question list. Neither needs outside help.

A partner earns their fee in three places. One is evaluation design: building a labelled set that reflects your real text, and setting an acceptance bar the business will stand behind. Another is the cost and governance model, where AI function usage, reservations and cross-region replication meet budgets and access rules. The third is disaster recovery itself: designing which datasets get replicated, running the drill, and handing over a runbook your own team can operate.

If you're weighing outside help, start with the BigQuery partners page and use the comparison tool to line up candidates. When you're ready to ask for proposals, the RFP template helps you state the workloads, the quality bar and the recovery objectives in one brief. Ask any firm to show how they measured quality on a past AI-function project, not just that they used the feature.

For more release coverage, see the MarTech Partners blog.

Sources

Frequently Asked Questions

What is optimized mode for AI.IF and AI.CLASSIFY in BigQuery?

It is a mode for the managed AI functions AI.IF and AI.CLASSIFY that aims to reduce LLM token consumption and query latency when you process large datasets. The 17 November 2026 release notes list it as generally available. Check Google's documentation for how to switch it on and which queries qualify.

What does AI.RANK do in BigQuery?

AI.RANK reorders the rows of a table by how well the text in a column you specify answers a search query. Google lists it as generally available in the 17 November 2026 notes. It suits cases where you already have a candidate set and want it ordered by relevance.

Can I manage BigQuery cross-region dataset replication through the REST API?

Yes. The 17 November 2026 notes say datasets.insert, datasets.patch, datasets.update and datasets.get can configure and monitor cross-region dataset replication, and attach or detach failover reservations for managed disaster recovery. That makes the setup scriptable and reviewable in code.

Does optimized mode change the answers I get from AI.CLASSIFY?

The notes say the Job information tab in the Query results pane shows information about the relative quality of the optimized query results, which implies results can differ from a standard run. Compare against a labelled sample before relying on it. Google's documentation describes the details.